While the fluency and coherence of Large Language Models (LLMs) in text generation have seen significant improvements, their competency in generating appropriate expressions of uncertainty remains limited.Using a multilingual closed-book QA task and GPT-3.5, we explore how well LLMs are calibrated and express certainty across a diverse set of languages, including low-resource settings. Our results reveal strong performance in high-resource languages but a marked decline in performance in lower-resource languages. Across all, we observe an exaggerated expression of confidence in the model, which does not align with the correctness or likelihood of its responses. Our findings highlight the need for further research into accurate calibration of LLMs especially in a multilingual setting.

The task of Visual Question Generation (VQG) is to generate human-like questions relevant to the given image. As VQG is an emerging research field, existing works tend to focus only on resource-rich language such as English due to the availability of datasets. In this paper, we propose the first Bengali Visual Question Generation task and develop a novel transformer-based encoder-decoder architecture that generates questions in Bengali when given an image. We propose multiple variants of models - (i) image-only: baseline model of generating questions from images without additional information, (ii) image-category and image-answer-category: guided VQG where we condition the model to generate questions based on the answer and the category of expected question. These models are trained and evaluated on the translated VQAv2.0 dataset. Our quantitative and qualitative results establish the first state of the art models for VQG task in Bengali and demonstrate that our models are capable of generating grammatically correct and relevant questions. Our quantitative results show that our image-cat model achieves a BLUE-1 score of 33.12 and BLEU-3 score of 7.56 which is the highest of the other two variants. We also perform a human evaluation to assess the quality of the generation tasks. Human evaluation suggests that image-cat model is capable of generating goal-driven and attribute-specific questions and also stays relevant to the corresponding image.

In Referring Expression Generation, model inputs are often composed of different representations, including the visual properties of the intended referent, its relative position and size, and the visual context. Yet, the extent to which this information influences the generation process of black-box neural models is largely unclear. We investigate the relative weighting of target, location, and context information in the attention components of a Transformer-based generation model. Our results show a general target bias, which, however, depends on the content of the generated expressions, pointing to interesting directions for future research.

Language-and-vision models have shown good performance in tasks such as image-caption matching and caption generation. However, it is challenging for such models to generate pragmatically correct captions, which adequately reflect what is happening in one image or several images. It is crucial to evaluate this behaviour to understand underlying reasons behind it. Here we explore to what extent contextual language-and-vision models are sensitive to different discourse, both textual and visual. In particular, we employ one of the multi-modal transformers (ViLBERT) and test if it can match descriptions and images, differentiating them from distractors of different degree of similarity that are sampled from different visual and textual contexts. We place our evaluation in the multi-sentence and multi-image setup, where images and sentences are expected to form a single narrative structure. We show that the model can distinguish different situations but it is not sensitive to differences within one narrative structure. We also show that performance depends on the task itself, for example, what modality remains unchanged in non-matching pairs or how similar non-matching pairs are to original pairs.

In any system that uses structured knowledge graph (KG) data as its underlying knowledge representation, KG-to-text generation is a useful tool for turning parts of the graph data into text that can be understood by humans. Recent work has shown that models that make use of pretraining on large amounts of text data can perform well on the KG-to-text task, even with relatively little training data on the specific graph-to-text task. In this paper, we build on this concept by using large language models to perform zero-shot generation based on nothing but the model’s understanding of the triple structure from what it can read. We show that ChatGPT achieves near state-of-the-art performance on some measures of the WebNLG 2020 challenge, but falls behind on others. Additionally, we compare factual, counter-factual and fictional statements, and show that there is a significant connection between what the LLM already knows about the data it is parsing and the quality of the output text.

The WebNLG task consists of mapping a knowledge graph to a text verbalising the con- tent of that graph. The 2017 WebNLG edi- tion required participating systems to gener- ate English text from a set of DBpedia triples, while the 2020 WebNLG+ challenge addition- ally included generation into Russian and se- mantic parsing of English and Russian texts. In contrast, WebNLG 2023 focuses on four under-resourced languages which are severely under-represented in research on text genera- tion, namely Breton, Irish, Maltese and Welsh. In addition, WebNLG 2023 once again includes Russian. In this paper, we present the organi- sation of the shared task (data, timeline, eval- uation), briefly describe the participating sys- tems and summarise results for participating systems.

We present our solution for the Russian RDF002 to-text generation task of the WebNLG Challenge 2023. We use the pretrained large language model named FRED-T5 (Zmitrovich et al., 2023) to finetune on the train dataset. Also, we propose several types of prompt and run experiments to analyze their effectiveness. Our submission achieves 0.373 TER on the test dataset, taking the first place according to the results of the automatic evaluation and outperforming the best result of the previous challenge by 0.025. The code of our solution is available at the following link: https://github.com/Ivan30003/webnlg_interno

This paper presents system descriptions of our submitted outputs for WebNLG Challenge 2023. We use mT5 in multi-task and multilingual settings to generate more fluent and reliable verbalizations of the given RDF triples. Furthermore, we introduce a partial decoding technique to produce more elaborate yet simplified outputs. Additionally, we demonstrate the significance of employing better translation systems in creating training data.

LLMs are great at tasks involving English which dominates in their training data. We explore their ability to address tasks involving languages that are severely under-represented in their training data. More specifically, we do this in the context of data-to-text generation for Irish, Maltese, Welsh and Breton. During the prompt-engineering phase we tested GPT-3.5 and~4 with a range of prompt types and formats on a small sample of example input/output pairs. We then fully evaluated the two most promising prompts in two scenarios: (i) direct generation into the under-resourced languages, and (ii) generation into English followed by translation into the under-resourced languages. We find that few-shot prompting works better for direct generation into under-resourced languages, but that the difference disappears when pivoting via English. The few-shot + translation system variants were submitted to the WebNLG 2023 shared task where they outperformed all other systems by substantial margins in all languages on all automatic metrics. We conclude that good performance can be achieved with state-of-the-art LLMs out-of-the box for under-resourced languages. However, best results (for Welsh) of BLEU 25.12, ChrF++ 0.55, and TER 0.64 are well below the lowest ranked English system at WebNLG’20 with BLEU 0.391, ChrF++ 0.579, and TER 0.564.

In this paper, we describe the submission of Dublin City University (DCU) and Trinity College Dublin (TCD) for the WebNLG 2023 shared task. We present a fully rule-based pipeline for generating Irish texts from DBpedia triple sets which comprises 4 components: triple lexicalisation, generation of noninflected Irish text, inflection generation, and post-processing.

This paper presents our submission to the WebNLG Challenge 2023 for generating text in several low-resource languages from RDF-triples. Our submission focuses on using machine translation for generating texts in Irish, Maltese, Welsh and Russian. While a simple and straightfoward approach, recent works have shown that using monolingual models for inference for multilingual tasks with the help of machine translation (translate-test) can out-perform multilingual models and training multilingual models on machine-translated data (translate-train) through careful tuning of the MT component. Our results show that this approach demonstrates competitive performance for this task even with limited data.

We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

Embedding representations of text are useful for downstream natural language processing tasks. Several universal sentence representation methods have been proposed with a particular focus on self-supervised pre-training approaches to leverage the vast quantities of unlabelled data. However, there are two challenges for generating rich embedding representations for a new document. 1) The latest rich embedding generators are based on very large costly transformer-based architectures. 2) The rich embedding representation of a new document is limited to only the information provided without access to any explicit contextual and temporal information that could potentially further enrich the representation. We propose efficient retrieval-augmented text embeddings (ERATE) that tackles the first issue and offers a method to tackle the second issue. To the best of our knowledge, we are the first to incorporate retrieval to general purpose embeddings as a new paradigm, which we apply to the semantic similarity tasks of SentEval. Despite not reaching state-of-the-art performance, ERATE offers key insights that encourages future work into investigating the potential of retrieval-based embeddings.

Low-quality data can cause downstream problems in high-stakes applications. Data-centric approach emphasizes on improving dataset quality to enhance model performance. High-quality datasets are needed for general-purpose Large Language Models (LLMs) training, as well as for domain-specific models, which are usually small in size as it is costly to engage a large number of domain experts for their creation. Thus, it is vital to ensure high-quality domain-specific training data. In this paper, we propose a framework for enhancing the data quality of original datasets. (Code and dataset are available at https://github.com/IvaBojic/framework). We applied the proposed framework to four biomedical datasets and showed relative improvement of up to 33%/40% for fine-tuning of retrieval/reader models on the BioASQ dataset when using back translation to enhance the original dataset quality.

Context-aware translation can be achieved by processing a concatenation of consecutive sentences with the standard Transformer architecture. This paper investigates the intuitive idea of providing the model with explicit information about the position of the sentences contained in the concatenation window. We compare various methods to encode sentence positions into token representations, including novel methods. Our results show that the Transformer benefits from certain sentence position encoding methods on English to Russian translation, if trained with a context-discounted loss. However, the same benefits are not observed on English to German. Further empirical efforts are necessary to define the conditions under which the proposed approach is beneficial.

Pretrained language models (PLMs) on domain-specific data have been proven to be effective for in-domain natural language processing (NLP) tasks. Our work aimed to develop a language model which can be effective for the NLP tasks with the data from diverse social media platforms. We pretrained a language model on Twitter and Reddit posts in English consisting of 929M sequence blocks for 112K steps. We benchmarked our model and 3 transformer-based models—BERT, BERTweet, and RoBERTa on 40 social media text classification tasks. The results showed that although our model did not perform the best on all of the tasks, it outperformed the baseline model—BERT on most of the tasks, which illustrates the effectiveness of our model. Also, our work provides some insights of how to improve the efficiency of training PLMs.

We introduce two simple randomized variants of byte pair encoding (BPE) and explore whether randomizing the selection of merge operations substantially affects a downstream machine translation task. We focus on translation into morphologically rich languages, hypothesizing that this task may show sensitivity to the method of choosing subwords. Analysis using a Bayesian linear model indicates that one variant performs nearly indistinguishably compared to standard BPE while the other degrades performance less than we anticipated. We conclude that although standard BPE is widely used, there exists an interesting universe of potential variations on it worth investigating. Our code is available at: https://github.com/bltlab/random-bpe.

In recent years, there has been growing interest in the field of abstractive text summarization with focused contributions in relevant model architectures, datasets, and evaluation metrics. Despite notable research advances, previous works have identified certain limitations concerning the quality of datasets and the effectiveness of evaluation techniques for generated summaries. In this context, we examine these limitations further with the help of three quality measures, namely, Information Coverage, Entity Hallucination, and Summarization Complexity. As a part of this work, we investigate two widely used datasets (XSUM and CNNDM) and three existing models (BART, PEGASUS, and BRIO) and report our findings. Some key insights are: 1) Cumulative ROUGE score is an inappropriate evaluation measure since few high-scoring samples dominate the overall performance, 2) Existing summarization models have limited capability for information coverage and hallucinate to generate factual information, and 3) Compared to the model generated summaries, the reference summaries have lowest information coverage and highest entity hallucinations reiterating the need of new and better reference summaries.

The number of scientific publications in the biomedical domain is continuously increasing with time. An efficient system for indexing these publications is required to make the information accessible according to the user’s information needs. Task 10a of the BioASQ challenge aims to classify PubMed articles according to the MeSH ontology so that new publications can be grouped with similar preexisting publications in the field without the assistance of time-consuming and costly annotations by human annotators. In this work, we use Graph Neural Network (GNN) in the link prediction setting to exploit potential graph-structured information present in the dataset which could otherwise be neglected by transformer-based models. Additionally, we provide error analysis and a plausible reason for the substandard performance achieved by GNN.

In this paper, we describe how we unearthed some fundamental problems while building an analogy dataset modelled on BATS (Gladkova et al., 2016) to evaluate historical Irish embeddings on their ability to detect orthographic, morphological and semantic similarity.performance of our models in the analogy task was extremely poor regardless of the architecture, hyperparameters and evaluation metrics, while the qualitative evaluation revealed positive tendencies. argue that low agreement between field experts on fundamental lexical and orthographic issues, and the lack of a unified editorial standard in available resources make it impossible to build reliable evaluation datasets for computational models and obtain interpretable results. We emphasise the need for such a standard, particularly for NLP applications, and prompt Celticists and historical linguists to engage in further discussion. We would also like to draw NLP scholars’ attention to the role of data and its (extra)linguistic properties in testing new models, technologies and evaluation scenarios.

Recent research has demonstrated impressive generalization capabilities of several Knowledge Base Question Answering (KBQA) models on the GrailQA dataset. We inspect whether these models can generalize to other datasets in a zero-shot setting. We notice a significant drop in performance and investigate the causes for the same. We observe that the models are dependent not only on the structural complexity of the questions, but also on the linguistic styles of framing a question. Specifically, the linguistic dimensions corresponding to explicitness, readability, coherence, and grammaticality have a significant impact on the performance of state-of-the-art KBQA models. Overall our results showcase the brittleness of such models and the need for creating generalizable systems.

Active learning has been widely used in the task of text classification for its ability to select the most valuable samples to annotate while improving the model performance. However, the efficiency of active learning in multi-label text classification tasks has been under-explored due to the label imbalanceness problem. In this paper, we conduct an empirical study of active learning on multi-label text classification and evaluate the efficiency of five active learning strategies on six multi-label text classification tasks. The experiments show that some strategies in the single-label setting especially in imbalanced datasets.

We probe structural and discourse aspects of coreferential relationships in a fine-tuned Dutch BERT event coreference model. Previous research has suggested that no such knowledge is encoded in BERT-based models and the classification of coreferential relationships ultimately rests on outward lexical similarity. While we show that BERT can encode a (very) limited number of these discourse aspects (thus disproving assumptions in earlier research), we also note that knowledge of many structural features of coreferential relationships is absent from the encodings generated by the fine-tuned BERT model.

Despite recent successes in language models, their ability to represent numbers is insufficient. Humans conceptualize numbers based on their magnitudes, effectively projecting them on a number line; whereas subword tokenization fails to explicitly capture magnitude by splitting numbers into arbitrary chunks. To alleviate this shortcoming, alternative approaches have been proposed that modify numbers at various stages of the language modeling pipeline. These methods change either the (1) notation in which numbers are written (eg scientific vs decimal), the (2) vocabulary used to represent numbers or the entire (3) architecture of the underlying language model, to directly regress to a desired number. Previous work suggests that architectural change helps achieve state-of-the-art on number estimation but we find an insightful ablation - changing the model”s vocabulary instead (eg introduce a new token for numbers in range 10-100) is a far better trade-off. In the context of masked number prediction, a carefully designed tokenization scheme is both the simplest to implement and sufficient, ie with similar performance to the state-of-the-art approach that requires making significant architectural changes. Finally, we report similar trends on the downstream task of numerical fact estimation (for Fermi Problems) and discuss reasons behind our findings.

Conventional techniques for detecting online hate speech rely on the availability of a sufficient number of annotated instances, which can be costly and time consuming. For this reason, zero-shot or few-shot detection can offer an attractive alternative. In this paper, we explore a zero-shot detection approach based on natural language inference (NLI) models. Since the performance of the models in this approach depends heavily on the choice of a hypothesis, our goal is to determine which factors affect the quality of detection. We conducted a set of experiments with three NLI models and four hate speech datasets. We demonstrate that a zero-shot NLI-based approach is competitive with approaches that require supervised learning, yet they are highly sensitive to the choice of hypothesis. In addition, our experiments indicate that the results for a set of hypotheses on different model-data pairs are positively correlated, and that the correlation is higher for different datasets when using the same model than it is for different models when using the same dataset. These results suggest that if we find a hypothesis that works well for a specific model and domain or for a specific type of hate speech, we can use that hypothesis with the same model also within a different domain. While, another model might require different suitable hypotheses in order to demonstrate high performance.

British Sign Language (BSL) is a complex language with its own vocabulary and grammatical structure, separate from English. Despite its long-standing and widespread use by Deaf communities within the UK, thus far, there have been no effective tools for translating written English into BSL. This overt lack of available resources made learning the language highly inaccessible for most people, exacerbating the communication barrier between hearing and Deaf individuals. This paper introduces a rule-based translation system, designed with the ambitious aim of creating the first web application that is not only able to translate sentences in written English into a BSL video output, but can also serve as a learning aid to empower the development of BSL proficiency.

This paper presents the TFIDF based LSTM and Hierarchical Attention Networks (HAN) for code-mixed abusive comment detection and sentiment analysis for Dravidian languages. The traditional TF-IDF-based techniques have out- performed the Hierarchical Attention models in both the sentiment analysis and abusive language detection tasks. The Tulu sentiment analysis system demonstrated better performance for the Positive and Neutral classes, whereas the Tamil sentiment analysis system exhibited lower performance overall. This highlights the need for more balanced datasets and additional research to enhance the accuracy of sentiment analysis in the Tamil language. In terms of abusive language detection, the TF-IDF-LSTM models generally outperformed the Hierarchical Attention models. However, the mixed models displayed better performance for specific classes such as “Homophobia” and “Xenophobia.” This implies that considering both code-mixed and original script data can offer a different perspective for research in social media analysis.

Social media has become a vital platform for personal communication. Its widespread use as a primary means of public communication offers an exciting opportunity for early detection and management of mental health issues. People often share their emotions on social media, but understanding the true depth of their feelings can be challenging. Depression, a prevalent problem among young people, is of particular concern due to its link with rising suicide rates. Identifying depression levels in social media texts is crucial for timely support and prevention of negative outcomes. However, it’s a complex task because human emotions are dynamic and can change significantly over time. The DepSign-LT-EDI@RANLP 2023 shared task aims to classify social media text into three depression levels: “Not Depressed,” “Moderately Depressed,” and “Severely Depressed.” This overview covers task details, dataset, methodologies used, and results analysis. Roberta-based models emerged as top performers, with the best result achieving an impressive macro F1-score of 0.584 among 31 participating teams.

This paper manifest the overview of the shared task on Speech Recognition for Vulnerable individuals in Tamil(LT-EDI-ACL2023). Task is provided with an Tamil dataset, which is collected from elderly people of three different genders, male, female and transgender. The audio samples were recorded from the public locations like hospitals, markets, vegetable shop, etc. The dataset is released in two phase, training and testing phase. The partcipants were asked to use different models and methods to handle audio signals and submit the result as transcription of the test samples given. The result submitted by the participants was evaluated using WER (Word Error Rate). The participants used the transformer-based model for automatic speech recognition. The results and different pre-trained transformer based models used by the participants is discussed in this overview paper.

We present an overview of the second shared task on homophobia/transphobia Detection in social media comments. Given a comment, a system must predict whether or not it contains any form of homophobia/transphobia. The shared task included five languages: English, Spanish, Tamil, Hindi, and Malayalam. The data was given for two tasks. Task A was given three labels, and Task B fine-grained seven labels. In total, 75 teams enrolled for the shared task in Codalab. For task A, 12 teams submitted systems for English, eight teams for Tamil, eight teams for Spanish, and seven teams for Hindi. For task B, nine teams submitted for English, 7 teams for Tamil, 6 teams for Malayalam. We present and analyze all submissions in this paper.

Hope serves as a powerful driving force that encourages individuals to persevere in the face of the unpredictable nature of human existence. It instills motivation within us to remain steadfast in our pursuit of important goals, regardless of the uncertainties that lie ahead. In today’s digital age, platforms such as Facebook, Twitter, Instagram, and YouTube have emerged as prominent social media outlets where people freely express their views and opinions. These platforms have also become crucial for marginalized individuals seeking online assistance and support[1][2][3]. The outbreak of the pandemic has exacerbated people’s fears around the world, as they grapple with the possibility of losing loved ones and the lack of access to essential services such as schools, hospitals, and mental health facilities.

Part of Speech (POS) taggers for Swedish routinely fail for the third person gender-neutral pronoun “hen”, despite the fact that it has been a well-established part of the Swedish language since at least 2014. In addition to simply being a form of gender bias, this failure can have negative effects on other tasks relying on POS information. We demonstrate the usefulness of semi-synthetic augmented datasets in a case study, retraining a POS tagger to correctly recognize “hen” as a personal pronoun. We evaluate our retrained models for both tag accuracy and on a downstream task (dependency parsing) in a classicial NLP pipeline. Our results show that adding such data works to correct for the disparity in performance. The accuracy rate for identifying “hen” as a pronoun can be brought up to acceptable levels with only minor adjustments to the tagger’s vocabulary files. Performance parity to gendered pronouns can be reached after retraining with only a few hundred examples. This increase in POS tag accuracy also results in improvements for dependency parsing sentences containing hen.

Machine translation, and more specifically neural machine translation (NMT), have been proven to be subject to gender bias in recent years. Many studies have focused on evaluating and reducing this phenomenon, mainly through the analysis of occupational nouns’ translation for the same type of language combinations. In this paper, we reproduce a similar test set than in previous studies to investigate the influence of stereotypes and language combinations’ nature (formed with English, French and Italian) on gender bias occurrence in NMT. Similarly to previous studies, we confirm stereotypes as a major source of gender bias, especially in female contexts, while observing bias even in language combinations traditionally less examined.

Our research in Natural Language Processing (NLP) aims to detect hate speech comments specifically targeted at the LGBTQ+ community within the YouTube platform shared task conducted by LTEDI workshop. The dataset provided by the organizers exhibited a high degree of class imbalance, and to mitigate this, we employed NLPAUG, a data augmentation library. We employed several classification methods and reported the results using recall, precision, and F1-score metrics. The classification models discussed in this paper include a Bidirectional Long Short-Term Memory (BiLSTM) model trained with Word2Vec embeddings, a BiLSTM model trained with Twitter GloVe embeddings, transformer models such as BERT, DistiBERT, RoBERTa, and XLM-RoBERTa, all of which were trained and fine-tuned. We achieved a weighted F1-score of 0.699 on the test data and secured fifth place in task B with 7 classes for the English language.

Homophobia and Transphobia is defined as hatred or discomfort towards Gay, Lesbian, Transgender or Bisexual people. With the increase in social media, communication has become free and easy. This also means that people can also express hatred and discomfort towards others. Studies have shown that these can cause mental health issues. Thus detection and masking/removal of these comments from the social media platforms can help with understanding and improving the mental health of LGBTQ+ people. In this paper, GPT2 is used to detect homophobic and/or transphobic comments in social media comments. The comments used in this paper are from five (English, Spanish, Tamil, Malayalam and Hindi) languages. The results show that detecting comments in English language is easier when compared to the other languages.

Depression is a prevalent mental illness characterized by feelings of sadness and a lack of interest in daily activities. Early detection of depression is crucial to prevent severe consequences, making it essential to observe and treat the condition at its onset. At ACL-2022, the DepSign-LT-EDI project aimed to identify signs of depression in individuals based on their social media posts, where people often share their emotions and feelings. Using social media postings in English, the system categorized depression signs into three labels: “not depressed,” “moderately depressed,” and “severely depressed.” To achieve this, our team has applied MentalRoBERTa, a model trained on big data of mental health. The test results indicated a macro F1-score of 0.439, ranking the fourth in the shared task.

Depression, a widespread mental health disorder, affects a significant portion of the global population. Timely identification and intervention play a crucial role in ensuring effective treatment and support. Therefore, this research paper proposes a fine-tuned RoBERTa-based model for identifying depression in social media posts. In addition to the proposed model, Sentence-BERT is employed to encode social media posts into vector representations. These encoded vectors are then utilized in eight different popular classical machine learning models. The proposed fine-tuned RoBERTa model achieved a best macro F1-score of 0.55 for the development dataset and a comparable score of 0.41 for the testing dataset. Additionally, combining Sentence-BERT with Naive Bayes (S-BERT + NB) outperformed the fine-tuned RoBERTa model, achieving a slightly higher macro F1-score of 0.42. This demonstrates the effectiveness of the approach in detecting depression from social media posts.

The rise of social media has led to a drastic surge in the dissemination of hostile and toxic content, fostering an alarming proliferation of hate speech, inflammatory remarks, and abusive language. The exponential growth of social media has facilitated the widespread circulation of hostile and toxic content, giving rise to an unprecedented influx of hate speech, incendiary language, and abusive rhetoric. The study utilized different techniques to represent the text data in a numerical format. Word embedding techniques aim to capture the semantic and syntactic information of the text data, which is essential in text classification tasks. The study utilized various techniques such as CNN, BERT, and N-gram to classify social media posts into depression and non-depression categories. Text classification tasks often rely on deep learning techniques such as Convolutional Neural Networks (CNN), while the BERT model, which is pre-trained, has shown exceptional performance in a range of natural language processing tasks. To assess the effectiveness of the suggested approaches, the research employed multiple metrics, including accuracy, precision, recall, and F1-score. The outcomes of the investigation indicate that the suggested techniques can identify symptoms of depression with an average accuracy rate of 56%.

This paper describes our multiclass classification system developed as part of the LT-EDI@RANLP-2023 shared task. We used a BERT-based language model to detect homophobic and transphobic content in social media comments across five language conditions: English, Spanish, Hindi, Malayalam, and Tamil. We retrained a transformer-based cross-language pretrained language model, XLM-RoBERTa, with spatially and temporally relevant social media language data. We found the inclusion of this spatio-temporal data improved the classification performance for all language and task conditions when compared with the baseline. We also retrained a subset of models with simulated script-mixed social media language data with varied performance. The results from the current study suggests that transformer-based language classification systems are sensitive to register-specific and language-specific retraining.

The paper describes the system for the 4th Shared task on “Detecting Signs of Depression from Social Media Text” at LT-EDI@RANLP 2023, which aimed to identify signs of depression on English social media texts. The solution comprised data cleaning and pre-processing, the use of additional data, a method to deal with data imbalance as well as fine-tuning of two transformer-based pre-trained language models, RoBERTa-Large and DeBERTa-V3-Large. Four model architectures were developed by leveraging different word embedding pooling methods, namely a RoBERTa-Large bidirectional GRU model using GRU pooling and three DeBERTa models using CLS pooling, mean pooling and max pooling, respectively. Although ensemble learning of DeBERTa’s pooling methods through majority voting was employed for better performance, the RoBERTa bidirectional GRU model managed to receive the 8th place out of 31 submissions with 0.42 Macro-F1 score.

Depression has become a common health problem impacting millions of individuals globally. Workplace stress and an unhealthy lifestyle have increased in recent years, leading to an increase in the number of people experiencing depressive symptoms. The spread of the epidemic has further exacerbated the problem. Early detection and precise prediction of depression are critical for early intervention and support for individuals at risk. However, due to the social stigma associated with the illness, many people are afraid to consult healthcare specialists, making early detection practically impossible. As a result, alternative strategies for depression prediction are being investigated, one of which is analyzing users’ social media posting behaviour. The organizers of LT-EDI@RANLP carried out a shared Task to encourage research in this area. Our team participated in the shared task and secured 21st rank with a macro F1 score 0f 0.36. This article provides a summary of the model presented in the shared task.

This work described the work of the team CIMAT-NLP on the Shared task of Detecting Signs of Depression from Social Media Text at LT-EDI@RANLP 2023, which consists of depression classification on three levels: “not depression”, “moderate” depression and “severe” depression on text from social media. In this work, we proposed two approaches: (1) a transformer model which can handle big text without truncation of its length, and (2) an ensemble of six binary Bag of Words. Our team placed fourth in the competition and found that models trained with our approaches could place second

Various biological, genetic, psychological or social factors that feature a target oriented life with chronic stress and frequent traumatic experiences, lead to pessimism and apathy. The massive scale of depression should be dealt with as a disease rather than a ‘phase’ that is neglected by the majority. However, not a lot of people are aware of depression and its impact. Depression is a serious issue that should be treated in the right way. Many people dealing with depression do not realize that they have it due to the lack of awareness. This paper aims to address this issue with a tool built on the blocks of machine learning. This model analyzes the public social media texts and detects the signs of depression under three labels namely “not depressed”, “moderately depressed”, and “severely depressed” with high accuracy. The ensembled model uses three learners namely Multi-Layered Perceptron, Support Vector Machine and Multinomial Naive Bayes Classifier. The distinctive feature in this model is that it uses Artificial Neural Networks, Classifiers, Regression and Voting Classifiers to compute the final result or output.

This paper describes our methodology adopted to participate in the multi-class classification task under the auspices of the Third Workshop on Language Technology for Equality, Diversity, Inclusion (LT-EDI) in the Recent Advances in Natural Language Processing (RANLP) 2023 conference. The overall objective was to employ ML algorithms to detect signs of depression in English social media content, classifying each post into one of three categories: no depression, moderate depression, and severe depression. To accomplish this we utilized generative pretrained transformers (GPTs), leveraging the full-scale OpenAI API. Our strategy incorporated prompt engineering for zero-shot and few-shot learning scenarios with ChatGPT and fine-tuning a GPT-3 model. The latter approach yielded the best results which allowed us to outperform our benchmark XGBoost classifier based on character-level features on the dev set and score a macro F1 score of 0.419 on the final blind test set.

HOPE speeches convey uplifting and motivating messages that help enhance mental health and general well-being. Hope speech detection has gained popularity in the field of natural language processing as it gives people the motivation they need to face challenges in life. The momentum behind this technology has been fueled by the demand for encouraging reinforcement online. In this paper, a deep learning approach is proposed in which four different word embedding techniques are used in combination with capsule networks, and a comparative analysis is performed to obtain results. Oversampling is used to address class imbalance problem. The dataset used in this paper is a part of the LT-EDI RANLP 2023 Hope Speech Detection shared task. The approach proposed in this paper achieved a Macro Average F1 score of 0.49 and 0.62 in English and Hindi-English code mix test data, which secured 2nd and 3rd rank respectively in the above mentioned share task.

Depression is a severe mental health disorder characterized by persistent feelings of sadness and anxiety, a decline in cognitive functioning resulting in drastic changes in a human’s psychological and physical well-being. However, depression is curable completely when treated at a suitable time and treatment resulting in the rejuvenation of an individual. The objective of this paper is to devise a technique for detecting signs of depression from English social media comments as well as classifying them based on their intensity into severe, moderate, and not depressed categories. The paper illustrates three approaches that are developed when working toward the problem. Of these approaches, the BERT model proved to be the most suitable model with an F1 macro score of 0.407, which gave us the 11th rank overall.

An Automatic Speech Recognition systems for Tamil are designed to convert spoken lan- guage or speech signals into written Tamil text. Seniors go to banks, clinics and authoritative workplaces to address their regular necessities. A lot of older people are not aware of the use of the facilities available in public places or office. They need a person to help them. Like- wise, transgender people are deprived of pri- mary education because of social stigma, so speaking is the only way to help them meet their needs. In order to build speech enabled systems, spontaneous speech data is collected from seniors and transgender people who are deprived of using these facilities for their own benefit. The proposed system is developed with pretraind models are IIT Madras transformer ASR model and akashsivanandan/wav2vec2- large-xls-r-300m-tamil model. Both pretrained models are used to evaluate the test speech ut- terances, and obtainted the WER as 37.7144% and 40.55% respectively.

Submission of the paper for the result submitted in Shared Task on Speech Recognition for Vulnerable Individuals in Tamil- LT-EDI-2023. The task is to develop an automatic speech recognition system for Tamil language. The dataset provided in the task is collected from the elderly people who converse in Tamil language. The proposed ASR system is designed with pre-trained model. The pre-trained model used in our system is fine-tuned with Tamil common voice dataset. The test data released from the task is given to the proposed system, now the transcriptions are generated for the test samples and the generated transcriptions is submitted to the task. The result submitted is evaluated by task, the evaluation metric used is Word Error Rate (WER). Our Proposed system attained a WER of 39.8091%.

The abusive content on social media networks is causing destructive effects on the mental well-being of online users. Homophobia refers to the fear, negative attitudes and feeling towards homosexuality. Transphobia refer to negative attitudes, hatred and prejudice towards transsexual people. Even though, some parts of the society have started to accept homosexuality and transsexuality, there are still a large set of the population opposing it. Hate speech targeting LGBTQ+ individuals, known as homophobia/transphobia speech, has become a growing concern. This has led to a toxic and unwelcoming environment for LGBTQ+ people on online platforms. This poses a significant societal issue, hindering the progress of equality, diversity, and inclusion. The identification of homophobic and transphobic comments on social media platforms plays a crucial role in creating a safer environment for all social media users. In order to accomplish this, we built a machine learning model using SGD and SVM classifier. Our approach yielded promising results, with a weighted F1-score of 0.95 on the English dataset and we secured 4th rank in this task.

This paper presents our ensembling solutions for detecting signs of depression in social media text, as part of the Shared Task at LT-EDI@RANLP 2023. By leveraging social media posts in English, the task involves the development of a system to accurately classify them as presenting signs of depression of one of three levels: “severe”, “moderate”, and “not depressed”. We verify the hypothesis that combining contextual information from a language model with local domain-specific features can improve the classifier’s performance. We do so by evaluating: (1) two global classifiers (support vector machine and logistic regression); (2) contextual information from language models; and (3) the ensembling results.

Many people may find motivation in their lives by spreading content on social media that is encouraging or hopeful. Creating an effective model that helps in accurately predicting the target class is a challenging task. The problem of Hope speech identification is dealt with in this work using machine learning and deep learning methods. This paper presents the description of the system submitted by our team(VEL) to the Hope Speech Detection for Equality, Diversity, and Inclusion(HSD-EDI) LT-EDI-RANLP 2023 shared task for the Bulgarian language. The main goal of this shared task is to identify the given text into the Hope speech or Non-Hope speech category. The proposed method used the H2O deep learning model with MPNet embeddings and achieved the second rank for the Bulgarian language with the Macro F1 score of 0.69.

Detecting transphobia, homophobia, and various other forms of hate speech is difficult. Signals can vary depending on factors such as language, culture, geographical region, and the particular online platform. Here, we present a joint multilingual (M-L) and language-specific (L-S) approach to homophobia and transphobic hate speech detection (HSD). M-L models are needed to catch words, phrases, and concepts that are less common or missing in a particular language and subsequently overlooked by L-S models. Nonetheless, L-S models are better situated to understand the cultural and linguistic context of the users who typically write in a particular language. Here we construct a simple and successful way to merge the M-L and L-S approaches through simple weight interpolation in such a way that is interpretable and data-driven. We demonstrate our system on task A of the “Shared Task on Homophobia/Transphobia Detection in social media comments” dataset for homophobia and transphobic HSD. Our system achieves the best results in three of five languages and achieves a 0.997 macro average F1-score on Malayalam texts.

Depression is debilitating, and not uncommon. Indeed, studies of excessive social media users show correlations with depression, ADHD, and other mental health concerns. Given that there is a large number of people with excessive social media usage, then there is a significant population of potentially undiagnosed users and posts that they create. In this paper, we propose a depression detection system using a semi-supervised learning technique. Namely, we use a trained model to classify a large number of unlabelled social media posts from Reddit, then use these generated labels to train a more powerful classifier. We demonstrate our framework on Detecting Signs of Depression from Social Media Text - LT-EDI@RANLP 2023 shared task, where our framework ranks 3rd overall.

Depression is a mental fitness disorder from persistent reactions of unhappiness, void, and a deficit of interest in activities. It can influence differing facets of one’s life, containing their hopes, sympathy, and nature. Depression can stem from a sort of determinant, in the way that ancestral willingness, life occurrences, and social circumstances. In current years, the influence of social media on mental fitness has become an increasing concern. Excessive use of social media and the negative facets that guide it, can exacerbate or cause impressions of distress. The nonstop exposure to cautiously curated lives, social comparison, cyberbullying, and the pressure to meet unreal standards can impact an individual’s pride, social connections, and overall well-being. We participated in the shared task at DepSignLT-EDI@RANLP 2023 and have proposed a model that identifies the levels of depression from social media text using the data set shared for the task. Different transformer models like ALBERT and RoBERTa are used by the proposed model for implementing the task. The macro F1 score obtained by ALBERT model and RoBERTa model are 0.258 and 0.143 respectively.

This paper centers on utilizing Automatic Speech Recognition (ASR) for defenseless old-aged and transgender people in Tamil. The Amrrs/wav2vec2-large-xlsr-53-tamil show accomplishes a Word Error Rate (WER) of 40%. By leveraging this demonstration, ASR innovation upgrades availability and inclusivity, helping those with discourse impedances, hearing impedances, and cognitive inabilities. Assist refinements are vital to diminish error and move forward the client involvement. This inquiry emphasizes the significance of ASR, particularly the Amrrs/wav2vec2-large-xlsr-53-tamil show, in encouraging successful communication and availability for defenseless populaces in Tamil.

Hope speech embodies optimistic and uplifting sentiments, aiming to inspire individuals to maintain faith in positive progress and actively contribute to a better future. In this article, we outline the model presented by our team, VTUBGM, for the shared task “Hope Speech Detection for Equality, Diversity, and Inclusion” at LT-EDI-RANLP 2023. This task entails classifying YouTube comments, which is a classification problem at the comment level. The task was conducted in four different languages: Bulgarian, English, Hindi, and Spanish. VTUBGM submitted a model developed through layered differential training of the ULMFit model. As a result, a macro F1 score of 0.48 was obtained and ranked 3rd in the competition.

Hope speech analysis refers to the examination and evaluation of speeches or messages that aim to instill hope, inspire optimism, and motivate individuals or communities. It involves analyzing the content, language, rhetorical devices, and delivery techniques used in a speech to understand how it conveys hope and its potential impact on the audience. The objective of this study is to classify the given text comments as Hope Speech or Not Hope Speech. The provided dataset consists of YouTube comments in four languages: English, Hindi, Spanish, Bulgarian; with pre-defined classifications. Our approach involved pre-processing the dataset and using the TF-IDF (Term Frequency-Inverse Document Frequency) method.

DepSign–LT–EDI@RANLP–2023 is a dedicated task that addresses the crucial issue of identifying indications of depression in individuals through their social media posts, which serve as a platform for expressing their emotions and sentiments. The primary objective revolves around accurately classifying the signs of depression into three distinct categories: “not depressed,” “moderately depressed,” and “severely depressed.” Our study entailed the utilization of machine learning algorithms, coupled with a diverse range of features such as sentence embeddings, TF-IDF, and Bag-of- Words. Remarkably, the adoption of hybrid models yielded promising outcomes, culminating in a 10^th rank achievement, supported by macro F1-Score of 0.408. This research underscores the effectiveness and potential of employing advanced text classification methodologies to discern and identify signs of depression within social media data. The findings hold implications for the development of mental health monitoring systems and support mechanisms, contributing to the well-being of individuals in need.

Our research aims to address the task of detecting homophobia and transphobia in social media code-mixed comments written in Spanish. Code-mixed text in social media often violates strict grammar rules and incorporates non-native scripts, posing challenges for identification. To tackle this problem, we perform pre-processing by removing unnecessary content and establishing a baseline for detecting homophobia and transphobia. Furthermore, we explore the effectiveness of various traditional machine-learning models with feature extraction and pre-trained transformer model techniques. Our best configurations achieve macro F1 scores of 0.84 on the test set and 0.82 on the development set for Spanish, demonstrating promising results in detecting instances of homophobia and transphobia in code-mixed comments.

As world population increases, more people are living to the age when depression or Major Depressive Disorder (MDD) commonly occurs. Consequently, the number of those who suffer from such disorders is rising. There is a pressing need for faster and reliable diagnosis methods. This paper proposes the method to analyse text input from social media posts of subjects to determine the severity class of depression. We have used the DistilBERT transformer to process these texts and classify the individuals across three severity labels - ‘not depression’, ‘moderate’ and ‘severe’. The results showed the macro F1-score of 0.437 when the model was trained for 5 epochs with a comparative performance across the labels.The team acquired 6th rank while the top team scored macro F1-score as 0.470. We hope that this system will support further research into the early identification of depression in individuals to promote effective medical research and related treatments.

Social media platforms have revolutionized the landscape of communication, providing individuals with an outlet to express their thoughts, emotions, and experiences openly. This paper focuses on the development of a model to determine whether individuals exhibit signs of depression based on their social media texts. With the aim of optimizing performance and accuracy, a Naive Bayes approach was chosen for the detection task.The Naive Bayes algorithm, a probabilistic classifier, was applied to extract features and classify the texts. The model leveraged linguistic patterns, sentiment analysis, and other relevant features to capture indicators of depression within the texts. Preprocessing techniques, including tokenization, stemming, and stop-word removal, were employed to enhance the quality of the input data.The performance of the Naive Bayes model was evaluated using standard metrics such as accuracy, precision, recall, and F1-score, it acheived a macro- avergaed F1 score of 0.263.

Social media platforms have become integral to our daily lives, facilitating instant sharing of thoughts and ideas. While these platforms often host inspiring, motivational, and positive content, the research community has recognized the significance of such messages by labeling them as “hope speech”. In light of this, we delve into the detection of hope speech on social media platforms. Specifically, we explore various transformer-based model setups for the LT-EDI shared task at RANLP 2023. We observe that the performance of the models varies across languages. Overall, the finetuned m-BERT model showcases the best performance among all the models across languages. Our models secured the first position in Bulgarian and Hindi languages and achieved the third position for the Spanish language in the respective task.

Hope is a cheerful and optimistic state of mind which has its basis in the expectation of positive outcomes. Hope speech reflects the same as they are positive words that can motivate and encourage a person to do better. Non-hope speech reflects the exact opposite. They are meant to ridicule or put down someone and affect the person negatively. The shared Task on Hope Speech Detection for Equality, Diversity, and Inclusion at LT-EDI - RANLP 2023 was created with data sets in English, Spanish, Bulgarian and Hindi. The purpose of this task is to classify human-generated comments on the platform, YouTube, as Hope speech or non-Hope speech. We employed multiple traditional models such as SVM (support vector machine), Random Forest classifier, Naive Bayes and Logistic Regression. Support Vector Machine gave the highest macro average F1 score of 0.49 for the training data set and a macro average F1 score of 0.50 for the test data set.

This submission presents our approach for depression detection in social media text. The methodology includes data collection, preprocessing - SMOTE, feature extraction/selection - TF-IDF and Glove, model development- SVM, CNN and Bi-LSTM, training, evaluation, optimisation, and validation. The proposed methodology aims to contribute to the accurate detection of depression.

The advent of social media platforms has revo- lutionized the way we interact, share, learn , ex- press and build our views and ideas. One major challenge of social media is hate speech. Homo- phobia and transphobia encompasses a range of negative attitudes and feelings towards people based on their sexual orientation or gender iden- tity. Homophobia refers to the fear, hatred, or prejudice against homosexuality, while trans- phobia involves discrimination against trans- gender individuals. Natural Language Process- ing can be used to identify homophobic and transphobic texts and help make social media a safer place. In this paper, we explore us- ing Support Vector Machine , Random Forest Classifier and Bert Model for homophobia and transphobia detection. The best model was a combination of LaBSE and SVM that achieved a weighted F1 score of 0.95.

In this paper, we delineate the strategy employed by our team, DeepLearningBrasil, which secured us the first place in the shared task DepSign-LT-EDI@RANLP-2023 with the advantage of 2.4%. The task was to classify social media texts into three distinct levels of depression - “not depressed,” “moderately depressed,” and “severely depressed.” Leveraging the power of the RoBERTa and DeBERTa models, we further pre-trained them on a collected Reddit dataset, specifically curated from mental health-related Reddit’s communities (Subreddits), leading to an enhanced understanding of nuanced mental health discourse. To address lengthy textual data, we introduced truncation techniques that retained the essence of the content by focusing on its beginnings and endings. Our model was robust against unbalanced data by incorporating sample weights into the loss. Cross-validation and ensemble techniques were then employed to combine our k-fold trained models, delivering an optimal solution. The accompanying code is made available for transparency and further development.

Hope plays a significant role in shaping human thoughts and actions and hope content has received limited attention in the realm of social media data analysis. The exploration of hope content helps to uncover the valuable insights into users’ aspirations, expectations, and emotional states. By delving into the analysis of hope content on social media platforms, researchers and analysts can gain a deeper understanding of how hope influences individuals’ behaviors, decisions, and overall well-being in the digital age. However, this area is rarely explored even for resource-high languages. To address the identification of hope text in social media platforms, this paper describes the models submitted by the team MUCS to “Hope Speech Detection for Equality, Diversity, and Inclusion (LT-EDI)” shared task organized at Recent Advances in Natural Language Processing (RANLP) - 2023. This shared task aims to classify a comment/post in English and code-mixed texts in three languages, namely, Bulgarian, Spanish, and Hindi into one of the two predefined categories, namely, “Hope speech” and “Non Hope speech”. Two models, namely: i) Hope_BERT - Linear Support Vector Classifier (LinearSVC) model trained by combining Bidirectional Encoder Representations from Transformers (BERT) embeddings and Term Frequency-Inverse Document Frequency (TF-IDF) of character n-grams with word boundary (char_wb) for English and ii) Hope_mBERT - LinearSVC model trained by combining Multilingual BERT (mBERT) embeddings and TF-IDF of char_wb for Bulgarian, Spanish, and Hindi code-mixed texts are proposed for the shared task to classify the given text into Hope or Non-Hope categories. The proposed models obtained 1st, 1st, 2nd, and 5th ranks for Spanish, Bulgarian, Hindi, and English texts respectively.

Homophobic/Transphobic (H/T) content includes hate speech, discrimination text, and abusive comments against Gay, Lesbian, Bisexual, Transgender, Queer, and Intersex (LGBTQ) individuals. With the increase in user generated text in social media, there has been an increase in code-mixed H/T content, which poses challenges for efficient analysis and detection of H/T content on social media. The complex nature of code-mixed text necessitates the development of advanced tools and techniques to effectively tackle this issue in social media platforms. To tackle this issue, in this paper, we - team MUCS, describe the transformer based models submitted to “Homophobia/Transphobia Detection in social media comments” shared task in Language Technology for Equality, Diversity and Inclusion (LT-EDI) at Recent Advances in Natural Language Processing (RANLP)-2023. The proposed methodology makes use of resampling the training data to handle the data imbalance and this resampled data is used to fine-tune the Multilingual Bidirectional Encoder Representations from Transformers (mBERT) models. These models obtained 11th, 5th, 3rd, 3rd, and 7th ranks for English, Tamil, Malayalam, Spanish, and Hindi respectively in Task A and 8th, 2nd, and 2nd ranks for English, Tamil, and Malayalam respectively in Task B.

Depression can lead to significant changes in individuals’ posts on social media which is a important task to identify. Automated techniques must be created for the identification task as manually analyzing the growing volume of social media data is time-consuming. To address the signs of depression posts on social media, in this paper, we - team MUCS, describe a Transfer Learning (TL) model and Machine Learning (ML) models submitted to “Detecting Signs of Depression from Social Media Text” shared task organised by DepSign-LT-EDI@RANLP-2023. The TL model is trained using raw text Bidirectional Encoder Representations from Transformers (BERT) and the ML model is trained using Term Frequency-Inverse Document Frequency (TF-IDF) features separately. Among these three models, the TL model performed better with a macro averaged F1-score of 0.361 and placed 20th rank in the shared task.

The goal of this study is to use machine learning approaches to detect depression indications in social media articles. Data gathering, pre-processing, feature extraction, model training, and performance evaluation are all aspects of the research. The collection consists of social media messages classified into three categories: not depressed, somewhat depressed, and severely depressed. The study contributes to the growing field of social media data-driven mental health analysis by stressing the use of feature extraction algorithms for obtaining relevant information from text data. The use of social media communications to detect depression has the potential to increase early intervention and help for people at risk. Several feature extraction approaches, such as TF-IDF, Count Vectorizer, and Hashing Vectorizer, are used to quantitatively represent textual data. These features are used to train and evaluate a wide range of machine learning models, including Logistic Regression, Random Forest, Decision Tree, Gaussian Naive Bayes, and Multinomial Naive Bayes. To assess the performance of the models, metrics such as accuracy, precision, recall, F1 score, and the confusion matrix are utilized. The Random Forest model with Count Vectorizer had the greatest accuracy on the development dataset, coming in at 92.99 percent. And with a macro F1-score of 0.362, we came in 19th position in the shared task. The findings show that machine learning is effective in detecting depression markers in social media articles.

The prevalence of depression is increasing globally, and there is a need for effective screening and detection tools. Social media platforms offer a rich source of data for mental health research. The paper aims to detect the signs of depression of a person from their social media postings wherein people share their feelings and emotions. The task is to create a system that, given social media posts in English, should classify the level of depression as ‘not depressed’, ‘moderately depressed’ or ‘severely depressed’. The paper presents the solution for the Shared Task on Detecting Signs of Depression from Social Media Text at LT-EDI@RANLP 2023. The proposed system aims to develop a machine learning model using machine learning algorithms like SVM, Random forest and Naive Bayes to detect signs of depression from social media text. The model is trained on a dataset of social media posts to detect the level of depression of the individuals as ‘not depressed’, ‘moderately depressed’ or ‘severely depressed’. The dataset is pre-processed to remove duplicates and irrelevant features, and then, feature engineering techniques is used to extract meaningful features from the text data. The model is trained on these features to classify the text into the three categories. The performance of the model is evaluated using metrics such as accuracy, precision, recall, and F1-score. The ensemble model is used to combine these algorithms which gives accuracy of 90.2% and the F1 score is 0.90. The results of the proposed approach could potentially aid in the early detection and prevention of depression for individuals who may be at risk.

Natural language processing (NLP) has recently gained relevance within financial institutions by providing highly valuable insights into companies and markets’ financial documents. However, the landscape of the financial domain presents extra challenges for NLP, due to the complexity of the texts and the use of specific terminology. Generalist language models tend to fall short in tasks specifically tailored for finance, even when using large language models (LLMs) with great natural language understanding and generative capabilities. This paper presents a study on LLM adaptation methods targeted at the financial domain and with high emphasis on financial sentiment analysis. To this purpose, two foundation models with less than 1.5B parameters have been adapted using a wide range of strategies. We show that through careful fine-tuning on both financial documents and instructions, these foundation models can be adapted to the target domain. Moreover, we observe that small LLMs have comparable performance to larger scale models, while being more efficient in terms of parameters and data. In addition to the models, we show how to generate artificial instructions through LLMs to augment the number of samples of the instruction dataset.

In this paper, we present ECL, a novel multimodal dataset containing the textual and numerical data from corporate 10K filings and associated binary bankruptcy labels. Furthermore, we develop and critically evaluate several classical and neural bankruptcy prediction models using this dataset. Our findings suggest that the information contained in each data modality is complementary for bankruptcy prediction. We also see that the binary bankruptcy prediction target does not enable our models to distinguish next year bankruptcy from an unhealthy financial situation resulting in bankruptcy in later years. Finally, we explore the use of LLMs in the context of our task. We show how GPT-based models can be used to extract meaningful summaries from the textual data but zero-shot bankruptcy prediction results are poor. All resources required to access and update the dataset or replicate our experiments are available on github.com/henriarnoUG/ECL.

The purpose of this paper is to construct a model for the generation of sophisticated headlines pertaining to stock price fluctuation articles, derived from the articles’ content. With respect to this headline generation objective, this paper solves three distinct tasks: in addition to the task of generating article headlines, two other tasks of extracting security names, and ascertaining the trajectory of stock prices, whether they are rising or declining. Regarding the headline generation task, we also revise the task as the model utilizes the outcomes of the security name extraction and rise/decline determination tasks, thereby for the purpose of preventing the inclusion of erroneous security names. We employed state-of-the-art pre-trained models from the field of natural language processing, fine-tuning these models for each task to enhance their precision. The dataset utilized for fine-tuning comprises a collection of articles delineating the rise and decline of stock prices. Consequently, we achieved remarkably high accuracy in the dual tasks of security name extraction and stock price rise or decline determination. For the headline generation task, a significant portion of the test data yielded fitting headlines.

Audit reports are a window to the financial health of a company and hence gauging coverage of various audit aspects in them is important. In this paper, we aim at determining an audit report’s coverage through classification of its sentences into multiple domain specific classes. In a weakly supervised setting, we employ a rule-based approach to automatically create training data for a BERT-based multi-label classifier. We then devise an ensemble to combine both the rule based and classifier approaches. Further, we employ two novel ways to improve the ensemble’s generalization: (i) through an active learning based approach and, (ii) through a LLM based review. We demonstrate that our proposed approaches outperform several baselines. We show utility of the proposed approaches to measure audit coverage on a large dataset of 2.8K audit reports.

Relation extraction (RE) is a crucial task in natural language processing (NLP) that aims to identify and classify relationships between entities mentioned in text. In the financial domain, relation extraction plays a vital role in extracting valuable information from financial documents, such as news articles, earnings reports, and company filings. This paper describes our solution to relation extraction on one such dataset REFinD. The dataset was released along with shared task as a part of the Fourth Workshop on Knowledge Discovery from Unstructured Data in Financial Services, co-located with SIGIR 2023. In this paper, we employed OpenAI models under the framework of in-context learning (ICL). We utilized two retrieval strategies to find top K relevant in-context learning demonstrations / examples from training data for a given test example. The first retrieval mechanism, we employed, is a learning-free dense retriever and the other system is a learning-based retriever. We were able to achieve 3rd rank overall. Our best F1-score is 0.718.

Assessing a company’s sustainable development goes beyond just financial metrics; the inclusion of environmental, social, and governance (ESG) factors is becoming increasingly vital. The ML-ESG shared task series seeks to pioneer discussions on news-driven ESG ratings, drawing inspiration from the MSCI ESG rating guidelines. In its second edition, ML-ESG-2 emphasizes impact type identification, offering datasets in four languages: Chinese, English, French, and Japanese. Of the 28 teams registered, 8 participated in the official evaluation. This paper presents a comprehensive overview of ML-ESG-2, detailing the dataset specifics and summarizing the performance outcomes of the participating teams.

The paper presents a concise summary of our work for the ML-ESG-2 shared task, exclusively on the Chinese and English datasets. ML-ESG-2 aims to ascertain the influence of news articles on corporations, specifically from an ESG perspective. To this end, we generally explored the capability of key information for impact identification and experimented with various techniques at different levels. For instance, we attempted to incorporate important information at the word level with TF-IDF, at the sentence level with TextRank, and at the document level with summarization. The final results reveal that the one with GPT-4 for summarisation yields the best predictions.

With the growing interest in Green Investing, Environmental, Social, and Governance (ESG) factors related to Institutions and financial entities has become extremely important for investors. While the classification of potential ESG factors is an important issue, identifying whether the factors positively or negatively impact the Institution is also a key aspect to consider while making evaluations for ESG scores. This paper presents our solution to identify ESG impact types in four languages (English, Chinese, Japanese, French) released as shared tasks during the FinNLP workshop at the IJCNLP-AACL-2023 conference. We use a combination of translation, masked language modeling, paraphrasing, and classification to solve this problem and use a generalized pipeline that performs well across all four languages. Our team ranked 1st in the Chinese and Japanese sub-tasks.

In this paper, we present our solutions to the ML-ESG-2 shared task which is co-located with the FinNLP workshop at IJCNLP-AACL-2023. The task proposes an objective of binary classification of ESG-related news based on what type of impact they can have on a company - Risk or Opportunity. We report the results of three systems, which ranked 2nd, 9th, and 10th in the final leaderboard for the English language, with the best solution achieving over 0.97 in F1 score.

This paper presents our findings in the ML-ESG-2 task, which focused on classifying a news snippet of various languages as “Risk” or “Opportunity” in the ESG (Environmental, Social, and Governance) context. We experimented with data augmentation and translation facilitated by Large Language Models (LLM). We found that augmenting the English dataset did not help to improve the performance. By fine-tuning RoBERTa models with the original data, we achieved the top position for the English and second place for the French task. In contrast, we could achieve comparable results on the French dataset by solely using the English translation, securing the third position for the French task with only marginal F1 differences to the second-place model.

In this paper, we describe our approach to the ML-ESG-2 shared task, co-located with the FinNLP workshop at IJCNLP-AACL-2023. The task aims at classifying news articles into categories reflecting either “Opportunity” or “Risk” from an ESG standpoint for companies. Our innovative methodology leverages two distinct systems for optimal text classification. In the initial phase, we engage in prompt engineering, working in conjunction with semantic similarity and using the Claude 2 LLM. Subsequently, we apply fine-tuning techniques to the Llama 2 and Dolly LLMs to enhance their performance. We report the results of five different approaches in this paper, with our top models ranking first in the French category and sixth in the English category.

In this paper, we discuss our (Team HHU’s) submission to the Multi-Lingual ESG Impact Type Identification task (ML-ESG-2). The goal of this task is to determine if an ESG-related news article represents an opportunity or a risk. We use an adapter-based framework in order to train multiple adapter modules which capture different parts of the knowledge present in the training data. Experimenting with various Adapter Fusion setups, we focus both on combining the ESG-aspect-specific knowledge, and on combining the language-specific-knowledge. Our results show that in both cases, it is possible to effectively compose the knowledge in order to improve the impact type determination.

In the evolving landscape of Environmental, Social, and Corporate Governance (ESG) impact assessment, the ML-ESG-2 shared task proposes identifying ESG impact types. To address this challenge, we present a comprehensive system leveraging ensemble learning techniques, capitalizing on early and late fusion approaches. Our approach employs four distinct models: mBERT, FlauBERT-base, ALBERT-base-v2, and a Multi-Layer Perceptron (MLP) incorporating Latent Semantic Analysis (LSA) and Term Frequency-Inverse Document Frequency (TF-IDF) features. Through extensive experimentation, we find that our early fusion ensemble approach, featuring the integration of LSA, TF-IDF, mBERT, FlauBERT-base, and ALBERT-base-v2, delivers the best performance. Our system offers a comprehensive ESG impact type identification solution, contributing to the responsible and sustainable decision-making processes vital in today’s financial and corporate governance landscape.

Relation extraction (RE) has achieved remarkable progress with the help of pre-trained language models. However, existing RE models are usually incapable of handling two situations: implicit expressions and long-tail relation classes, caused by language complexity and data sparsity. Further, these approaches and models are largely inaccessible to users who don’t have direct access to large language models (LLMs) and/or infrastructure for supervised training or fine-tuning. Rule-based systems also struggle with implicit expressions. Apart from this, Real world financial documents such as various 10-X reports (including 10-K, 10-Q, etc.) of publicly traded companies pose another challenge to rule-based systems in terms of longer and complex sentences. In this paper, we introduce a simple approach that consults training relations at test time through a nearest-neighbor search over dense vectors of lexico-syntactic patterns and provides a simple yet effective means to tackle the above issues. We evaluate our approach on REFinD and show that our method achieves state-of-the-art performance. We further show that it can provide a good start for human in the loop setup when a small number of annotations are available and it is also beneficial when domain experts can provide high quality patterns. Our code is available at 1.

Linguistic structures can implicitly imply diverse types of event relations that have been previously underexplored. For example, the sentence “John was cooking freshly made noodles for the family gathering” contains no explicit temporal indicators between the events, such as before. Despite this, it is easy for humans to conclude, based on syntax, that the noodles were made before John started cooking, and that the family gathering starts after John starts cooking. We introduce Linguistically enhanced Event TemporAl relation Framework (LEAF), a simple and effective approach to acquiring rich temporal knowledge of events from large-scale corpora. This method improves pre-trained language models by automatically extracting temporal relation knowledge from unannotated corpora using diverse temporal knowledge patterns. We begin by manually curating a comprehensive list of atomic patterns that imply temporal relations between events. These patterns involve event pairs in which one event is contained within the argument of the other. Using transitivity, we discover compositional patterns and assign labels to event pairs involving these patterns. Finally, we make language models learn the rich knowledge by pre-training with the acquired temporal relation supervision. Experiments show that our method outperforms or rivals previous models on two event relation datasets: MATRES and TB-Dense. Our approach is also simpler from past works and excels at identifying complex compositional event relations.

Recently, end-to-end trained models for multiple-choice commonsense question answering (QA) have delivered promising results. However, such question-answering systems cannot be directly applied in real-world scenarios where answer candidates are not provided. Hence, a new benchmark challenge set for open-ended commonsense reasoning (OpenCSR) has been recently released, which contains natural science questions without any predefined choices. On the OpenCSR challenge set, many questions require implicit multi-hop reasoning and have a large decision space, reflecting the difficult nature of this task. Existing work on OpenCSR sorely focuses on improving the retrieval process, which extracts relevant factual sentences from a textual knowledge base, leaving the important and non-trivial reasoning task outside the scope. In this work, we extend the scope to include a reasoner that constructs a question-dependent open knowledge graph based on retrieved supporting facts and employs a sequential subgraph reasoning process to predict the answer. The subgraph can be seen as a concise and compact graphical explanation of the prediction. Experiments on two OpenCSR datasets show that the proposed model achieves great performance on benchmark OpenCSR datasets.

In this paper, we describe M-FleNS, a multilingual flexible plug-and-play architecture designed to accommodate neural and symbolic modules, and initially instantiated with rule-based modules. We focus on using M-FleNS for the specific purpose of building new resources for Irish, a language currently under-represented in the NLP landscape. We present the general M-FleNS framework and how we use it to build an Irish Natural Language Generation system for verbalising part of the DBpedia ontology and building a multilayered dataset with rich linguistic annotations. Via automatic and human assessments of the output texts we show that with very limited resources we are able to create a system that reaches high levels of fluency and semantic accuracy, while having very low energy and memory requirements.

Large language models (LLMs) excel at processing and generating both text and code. However, LLMs have had limited applicability in grounded task-oriented dialogue as they are difficult to steer toward task objectives and fail to handle novel grounding. We present a modular and interpretable grounded dialogue system that addresses these shortcomings by composing LLMs with a symbolic planner and grounded code execution. Our system consists of a reader and planner: the reader leverages an LLM to convert partner utterances into executable code, calling functions that perform grounding. The translated code’s output is stored to track dialogue state, while a symbolic planner determines the next appropriate response. We evaluate our system’s performance on the demanding OneCommon dialogue task, involving collaborative reference resolution on abstract images of scattered dots. Our system substantially outperforms the previous state-of-the-art, including improving task success in human evaluations from 56% to 69% in the most challenging setting.

Frame semantic parsing is an important component of task-oriented dialogue systems. Current models rely on a significant amount training data to successfully identify the intent and slots in the user’s input utterance. This creates a significant barrier for adding new domains to virtual assistant capabilities, as creation of this data requires highly specialized NLP expertise. In this work we propose OpenFSP, a framework that allows for easy creation of new domains from a handful of simple labels that can be generated without specific NLP knowledge. Our approach relies on creating a small, but expressive, set of domain agnostic slot types that enables easy annotation of new domains. Given such annotation, a matching algorithm relying on sentence encoders predicts the intent and slots for domains defined by end-users. Experiments on the TopV2 dataset shows that our model trained on these simple labels have strong performance against supervised baselines.

Natural language programming automatically generates code based on a user’s text query. Recent solutions are either data-driven or natural language understanding (NLU)-driven. However, the data-driven synthesizer requires a large number of query-code pairs for training, which hinders its application to low-resource programming languages with growing domains whose functionality and grammar can be actively updated. NLU-driven synthesizers solve this problem, but their code generation is slow and their performance rapidly saturates in the presence of ever-increasing data. In this paper, we propose a circular training framework, Colead, which co-evolves both the data-driven synthesizer and the NLU-driven synthesizer to achieve high-quality code generation in the presence of data scarcity and domain growth. The NLU-driven synthesizer generates query-code pairs to update the data-driven synthesizer, which shares a part of its updated model to improve the NLU-driven synthesizers, enabling the co-evolution of both. Experiments show that Colead gives better results than the baselines in the presence of domain growth and data scarcity, and Colead consistently improves the performance of both data-driven and NLU-driven synthesizers over the co-evolvement.

Answering open-domain questions through unsupervised methods poses challenges for both machine-reading (MR) and language model (LM) -based approaches. The MR-based approach suffers from sparsity issues in extracted knowledge graphs (KGs), while the performance of the LM-based approach significantly depends on the quality of the retrieved context for questions. In this paper, we compare these approaches and propose a novel methodology that leverages directional predicate entailment (inference) to address these limitations. We use entailment graphs (EGs), with natural language predicates as nodes and entailment as edges, to enhance parsed KGs by inferring unseen assertions, effectively mitigating the sparsity problem in the MR-based approach. We also show EGs improve context retrieval for the LM-based approach. Additionally, we present a Boolean QA task, demonstrating that EGs exhibit comparable directional inference capabilities to large language models (LLMs). Our results highlight the importance of inference in open-domain QA and the improvements brought by leveraging EGs.

Modern dialog systems rely on Deep Learning to train transformer-based model architectures. These notoriously rely on large amounts of training data. However, the collection of conversational data is often a tedious and costly process. This is especially true for Task-Oriented Dialogs, where the system ought to help the user achieve specific tasks, such as making reservations. We investigate a controlled strategy for dialog synthesis. Our method generates utterances based on dialog annotations in a sequence-to-sequence manner. Besides exploring the viability of the approach itself, we also explore the effect of constrained beam search on the generation capabilities. Moreover, we analyze the effectiveness of the proposed method as a data augmentation by studying the impact the synthetic dialogs have on training dialog systems. We perform the experiments in multiple settings, simulating various amounts of ground-truth data. Our work shows that a controlled generation approach is a viable method to synthesize Task-Oriented Dialogs, that can in turn be used to train dialog systems. We were able to improve this process by utilizing constrained beam search.

Relation extraction (RE) is one of the tasks behind many relevant natural language processing (NLP) applications. Exploiting the information hidden in millions of scholarly articles by leveraging NLP, specifically RE, systems could benefit studies in specialized domains, e.g. biomedicine and biodiversity. Although deep learning (DL)-based methods have shown state-of-the-art performance in many NLP tasks including RE, DL for domain-specific RE systems has been hindered by the lack of expert-labeled datasets which are typically required to train such methods. In this paper, we take advantage of the zero-shot (i.e., not requiring any labeled data) capability of pattern-based methods for RE using a rule-based approach, combined with templates for natural language inference (NLI) transformer models. We present our hybrid method for RE that exploits the advantages of both methods, i.e., interpretability of rules and transferability of transformers. Evaluated on a corpus of biodiversity literature with annotated relations, our hybrid method demonstrated an improvement of up to 15 percentage points in recall and best performance over solely rule-based and transformer-based methods with F1-scores ranging from 89.61% to 96.75% for reproductive condition - temporal expression relations, and ranging from 85.39% to 89.90% for habitat - geographic location relations.

This paper explores a post-suppositional view on wh-questions and their answers with dynamic semantics. Inspired by Brasoveanu (2013); Charlow (2017); Bumford (2017), I propose a unified treatment of items like modified numerals, focus items, and wh-items: they (i) introduce a discourse referent (dref) in a non-deterministic way and (ii) impose definiteness tests (and additional tests) in a delayed, post-suppositional manner at the sentential / discourse level. Thus, with a question like “who smiled”, the (maximally informative) dref “the one(s) who smiled” is derived. A short answer like “Mary and Max” is considered another post-supposition-like, delayed test, checking whether the dref “the one(s) who smiled” is identical to (or includes) the sum “Mary⊕Max”. I analyze various question-related phenomena to see how far this proposal can go.

The paper extends a referentially transparent approach which has been successfully applied to the analysis of declarative quantified NPs to wh-phrases. This uses data from dialogical phenomena such as clarification interaction, anaphora, and incrementality as a guide to the design of wh-phrase meanings.

The sort of denotation a sentence is assigned is typically motivated by assumptions about the discourse function of sentences of that kind. For example, the notion that utterances which are functionally inquisitive (asking a question) suggest denotations which are semantically inquisitive (expressing the multiple licit responses to that question) is the cornerstone of interrogative meaning in frameworks like Alternative Semantics (Hamblin, 1973) and Inquisitive Semantics (Ciardelli et al., 2018). This paper argues that at least some kinds of questions systematically do not involve utterances with inquisitive content, based on novel observations of the Estonian discourse particle ega. Though ega is often labeled a ‘question particle’, it is used in both assertions and questions with sharply divergent discourse effects. I suggest that the relevant difference between assertive and questioning uses of ega is not semantic or sentence type-related, but rather reflects an interaction between a unified semantics for declaratives ega-sentences and different contexts of use. I then show that if we assume that ega presupposes that some aspect of the discourse context implicates the negation of ega’s prejacent, and that it occurs only in declarative sentences, we can derive its interpretation across a range of contexts: with the right combination of ingredients, we can ask questions with semantically uninquisitive sentences.

This paper investigates Farsi particle ‘mage’ in interrogatives, including both polar and constituent/Wh questions. I will show that ‘mage’ requires both contextual evidence and speaker’s prior belief in the sense that they contradict each other. While in polar questions (PQs) both types of bias can be straightforwardly expressed through the uttered proposition (cf. Mameni 2010), Wh-questions (WhQs) do not provide such a propositional object. To capture this difference, I propose Answerhood as the relevant notation that provides the necessary object source for ‘mage’ (inspired by Theiler 2021). The proposal establishes the felicity conditions and the meaning of ‘mage’ in relation to the (contextually) restricted answerhood in both polar and constituent questions.

This paper investigates the clausal embedding pattern of the Mandarin verb “xiang” (think) and reveals its internal anti-interrogative nature, with the possibility of “xiang Q” in certain cases. Through various stativity tests, I establish that the results are consistent with the generalization proposed by Özyıldız(2021), with “minor” deviations observed in the stativity of “xiang P” and the correlation with neg-raising. Additionally, I employ a semantic shift perspective to explain instances of neg-raising failure. Overall, this study sheds light on the unique characteristics of the verb “xiang” and contributes to a better cross-linguistic understanding of CP selection.

Prior studies on spoken languages have shown that indefinite and interrogative pronouns may be formally very similar. Our research aims to understand if sign languages exhibit this type of affinity. This paper presents an overview of the phenomenon and reports on the results of two studies: a cross-linguistic survey based on a sample of 30 sign languages and an empirical investigation conducted with three deaf consultants of Catalan Sign Language (LSC). Our research shows that, in sign languages, certain signs have both existential and interrogative readings and it identifies the environments that make existential interpretations available in LSC.

Terminology Alignment faces big challenges in NLP because of the dynamic nature of terms. Fortunately, over these last few years, Deep Learning models showed very good progress with several NLP tasks such as multilingual data resourcing, glossary building, terminology understanding. . . etc. In this work, we propose a new method for terminology alignment from a comparable corpus (Arabic/French languages) for the Algerian culture field. We aim to improve bilingual alignment based on contextual information of a term and to create a significant term bank i.e. a bilingual Arabic-French dictionary. We propose to create word embeddings for both Arabic and French languages using ELMO model focusing on contextual features of terms. Then, we mapp those embeddings using Seq2seq model. We use multilingual-BERT and All-MiniLM-L6 as baseline mod- els to compare terminology alignment results. Lastly we study the performance of these models by applying evaluation methods. Experimentation’s showed quite satisfying alignment results.

We propose a tool for the semi-automatic production of terminological databases, divided in the steps of corpus processing, terminology extraction, database population and management. With this tool it is possible to obtain a draft macrostructure (a lemma-list) and data for the microstructural level, such as grammatical (morphosyntactic patterns, gender, formation process) and semantic information (hypernyms, equivalence in another language, definitions and synonyms). In this paper we offer an overall description of the software and an evaluation of its performance, for which we used a linguistics corpus in English and Spanish.

Use of NLP Techniques in Translation by ChatGPT: Case Study Natural Language Processing (NLP) refers to a field of study within the domain of artificial intelligence (AI) and computational linguistics that focuses on the interaction between computers and human language. NLP seeks to develop computational models and algorithms capable of understanding, analyzing, and generating natural language text and speech (Brown et al., 1990). At its core, NLP aims to bridge the gap between human language and machine understanding by employing various techniques from linguistics, computer science, and statistics. It involves the application of linguistic and computational theories to process, interpret, and extract meaningful information from unstructured textual data (Bahdanau, Cho and Bengio, 2015). Researchers and practitioners in NLP employ diverse methodologies, including rule-based approaches, statistical models, machine learning techniques (such as neural networks), and more recently, deep learning architectures. These methodologies enable the development of robust algorithms that can learn from large-scale language data to improve the accuracy and effectiveness of language processing systems (Nilsson, 2010). NLP has numerous real-world applications across various domains, including information retrieval, virtual assistants, chatbots, social media analysis, sentiment monitoring, automated translation services, and healthcare, among others (kaynak). As the field continues to advance, NLP strives to overcome challenges such as understanding the nuances of human language, handling ambiguity, context sensitivity, and incorporating knowledge from diverse sources to enable machines to effectively communicate and interact with humans in a more natural and intuitive manner. Natural Language Processing (NLP) and translation are interconnected fields that share a symbiotic relationship, as NLP techniques and methodologies greatly contribute to the advancement and effectiveness of machine translation systems. NLP, a subfield of artificial intelligence (AI), focuses on the interaction between computers and human language. It encompasses a wide range of tasks, including text analysis, syntactic and semantic parsing, sentiment analysis, information extraction, and machine translation (Bahdanau, Cho and Bengio, 2014). NMT models employ deep learning architectures, such as recurrent neural networks (RNNs) and more specifically, long short-term memory (LSTM) networks, to learn the mapping between source and target language sentences. These models are trained on large-scale parallel corpora, consisting of aligned sentence pairs in different languages. The training process involves optimizing model parameters to minimize the discrepancy between predicted translations and human-generated translations (Wu et al., 2016) NLP techniques are crucial at various stages of machine translation. Preprocessing techniques, such as tokenization, sentence segmentation, and morphological analysis, help break down input text into meaningful linguistic units, making it easier for translation models to process and understand the content. Syntactic and semantic parsing techniques aid in capturing the structural and semantic relationships within sentences, improving the overall coherence and accuracy of translations. Furthermore, NLP-based methods are employed for handling specific translation challenges, such as handling idiomatic expressions, resolving lexical ambiguities, and addressing syntactic divergences between languages. For instance, statistical alignment models, based on NLP algorithms, enable the identification of correspondences between words or phrases in source and target languages, facilitating the generation of more accurate translations (kaynak). Several studies have demonstrated the effectiveness of NLP techniques in enhancing machine translation quality. For example, Bahdanau et al. (2015) introduced the attention mechanism, an NLP technique that enables NMT models to focus on relevant parts of the source sentence during translation. This attention mechanism significantly improved the translation quality of neural machine translation models. ChatGPT is a language model developed by OpenAI that utilizes the principles of Natural Language Processing (NLP) for various tasks, including translations. NLP is a field of artificial intelligence that focuses on the interaction between computers and human language. It encompasses a range of techniques and algorithms for processing, analyzing, and understanding natural language. When it comes to translation, NLP techniques can be applied to facilitate the conversion of text from one language to another. ChatGPT employs a sequence-to-sequence model, a type of neural network architecture commonly used in machine translation tasks. This model takes an input sequence in one language and generates a corresponding output sequence in the target language (OpenAI, 2023). The training process for ChatGPT involves exposing the model to large amounts of multilingual data, allowing it to learn patterns, syntax, and semantic relationships across different languages. This exposure enables the model to develop a general understanding of language structures and meanings, making it capable of performing translation tasks. To enhance translation quality, ChatGPT leverages the Transformer architecture, which has been highly successful in NLP tasks. Transformers utilize attention mechanisms, enabling the model to focus on different parts of the input sequence during the translation process. This attention mechanism allows the model to capture long-range dependencies and improve the overall coherence and accuracy of translations. Additionally, techniques such as subword tokenization, which divides words into smaller units, are commonly employed in NLP translation systems like ChatGPT. Subword tokenization helps handle out-of-vocabulary words and improves the model’s ability to handle rare or unknown words (GPT-4 Technical Report, 2023). As can be seen, there have been significant developments in artificial intelligence translations thanks to NLP. However, it is not possible to say that it has fully reached the quality of translation made by people. The only goal in artificial intelligence translations is to reach translations made by humans. In general, there are some fundamental differences between human and ChatGPT translations. Human-made translations and translations generated by ChatGPT (or similar language models) have several key differences (Kelly and Zetzsche, 2014; Koehn, 2010; Sutskever, Vinyals and Le, 2014; Costa-jussà and Fonollosa, 2018) Translation Quality: Human translators are capable of producing high-quality translations with a deep understanding of both the source and target languages. They can accurately capture the nuances, cultural references, idioms, and context of the original text. On the other hand, ChatGPT translations can sometimes be less accurate or may not fully grasp the intended meaning due to the limitations of the training data and the model’s inability to comprehend context in the same way a human can. While ChatGPT can provide reasonable translations, they may lack the finesse and precision of a human translator. Natural Language Processing: Human translators are skilled at processing and understanding natural language, taking into account the broader context, cultural implications, and the intended audience. They can adapt their translations to suit the target audience, tone, and purpose of the text. ChatGPT, although trained on a vast amount of text data, lacks the same level of natural language understanding. It often relies on pattern matching and statistical analysis to generate translations, which can result in less nuanced or contextually appropriate outputs. Subject Matter Expertise: Human translators often specialize in specific domains or subject areas, allowing them to have deep knowledge and understanding of technical or specialized terminology. They can accurately translate complex or industry-specific texts, ensuring the meaning is preserved. ChatGPT, while having access to a wide range of general knowledge, may struggle with domain-specific vocabulary or terminology, leading to inaccuracies or incorrect translations in specialized texts. Cultural Sensitivity: Human translators are well-versed in the cultural nuances of both the source and target languages. They can navigate potential pitfalls, adapt the translation to the cultural context, and avoid unintended offensive or inappropriate language choices. ChatGPT lacks this level of cultural sensitivity and may produce translations that are culturally tone-deaf or insensitive, as it lacks the ability to understand the subtleties and implications of language choices. Revision and Editing: Human translators go through an iterative process of revision and editing to refine their translations, ensuring accuracy, clarity, and quality. They can self-correct errors and refine their translations based on feedback or additional research. ChatGPT, while capable of generating translations, does not have the same ability to self-correct or improve based on feedback. It generates translations in a single pass, without the iterative refinement process that humans can employ. In summary, while ChatGPT can be a useful tool for generating translations, human-made translations generally outperform machine-generated translations in terms of quality, accuracy, contextuality, cultural sensitivity, and domain-specific expertise. In conclusion, NLP and machine translation are closely intertwined, with NLP providing essential tools, methodologies, and techniques that contribute to the development and improvement of machine translation systems. The integration of NLP methods has led to significant advancements in translation accuracy, fluency, and the ability to handle various linguistic complexities. As NLP continues to evolve, its impact on the field of machine translation is expected to grow, enabling the creation of more sophisticated and context-aware translation systems. On the basis of all this information, in this research, it is aimed to compare the translations from English to Turkish made by ChatGPT, one of the most advanced artificial intelligences, with the translations made by humans. In this context, an academic 1 page English text was chosen. The text was translated by both ChatGPT and a translator who is an academic in the field of translation and has 10 years of experience. Afterwards, two different translations were examined comparatively by 5 different translators who are experts in their fields. Semi-structured in-depth interviews were conducted with these translators. The aim of this study is to reveal the role of artificial intelligence tools in translation, which are increasing day by day and suggesting that there will be no need for language learning in the future. On the other hand, many translators argue that artificial intelligence and human translations can be understood. Therefore, if artificial intelligence is successful, there will be no profession called translator in the future. This research seems to be very useful in terms of shedding light on the future. The method of this research is semi-structured in-depth interview. References Bahdanau, D., Cho, K. and Bengio Y. (2015). Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations. Brown, P. F., Cocke, J., Pietra, S. A. D., Pietra, V. J. D., Jelinek, F., Lafferty, J. D., Mercer, R. L., and Roossin, P. S. A. (1990) statistical approach to machine translation. Computational linguistics 16, 2, 79–85. Costa-jussà, M. R., & Fonollosa, J. A. R. (2018). “An Overview of Neural Machine Translation.” IEEE Transactions on Neural Networks and Learning Systems. GPT-4 Technical Report (2023). https://arxiv.org/abs/2303.08774. Kelly, N. and Zetzsche, J. (2014). Found in Translation: How Language Shapes Our Lives and Transforms the World. USA: Penguin Book. Koehn, P. (2010). “Statistical Machine Translation.” Cambridge University Press. Nilsson, N. J. (2010). The Quest For AI- A History Of Ideas And Achievements. http://ai.standford.edu/ nilsson/. OpenAI (2023). https://openai.com/blog/chatgpt/. Sutskever, I., Vinyals, O., & Le, Q. V. (2014). “Sequence to Sequence Learning with Neural Networks.” Advances in Neural Information Processing Systems. Wu,Y. Schuster, M., Chen, Z., Le, Q. V. and Norouzi M. (2016). Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation. https://arxiv.org/pdf/1609.08144.pdf.

In the translation process, terminological resources are used to solve translation problems, so information on terminological equivalence is crucial to make the most appropriate choices in terms of translation equivalence. In the context of Machine translation, indeed, neural models have improved the state-of-the-art in Machine Translation considerably in recent years. However, they still underperform in domain-specific fields and in under-resourced languages. This is particularly evident in translating legal terminology for Arabic, where current Machine Translation outputs do not adhere to the contextual, linguistic, cultural, and terminological constraints posed by translating legal terms in Arabic. In this paper, we conduct a comparative qualitative evaluation and comprehensive error analysis on legal terminology translation in Phrase-Based Statistical Machine Translation and Neural Machine Translation in two translation language pairs: Arabic-English and Arabic-French. We propose an error typology taking the legal terminology translation from Arabic into account. We demonstrate our findings, highlighting the strengths and weaknesses of both approaches in the area of legal terminology translation for Arabic. We also introduce a multilingual gold standard dataset that we developed using our Arabic legal corpus. This dataset serves as a reliable benchmark and/or reference during the evaluation process to decide the degree of adequacy and fluency of the Phrase-Based Statistical Machine Translation and Neural Machine Translation systems.

Implementing technology in a modern-day classroom is an ongoing challenge. In this paper, we created a system for an automatic assessment of student answers using Latent Semantic Analysis (LSA) – a method with an underlying assumption that words with similar meanings will appear in the same contexts. The system will be used within digital lexical flash-cards for L2 vocabulary acquisition in a CLIL classroom. Results presented in this paper indicate that while LSA does well in creating semantic spaces for longer texts, it somewhat struggles with detecting topics in short texts. After obtaining LSA semantic spaces, answer accuracy was assessed by calculating the cosine similarity between a student’s answer and the golden standard. The answers were classified by accuracy using KNN, for both binary and multinomial classification. The results of KNN classification are as follows: precision P = 0.73, recall R = 1.00, F1 = 0.85 for binary classification, and P = 0.50, R = 0.47, F1 = 0.46 score for the multinomial classifier. The results are to be taken with a grain of salt, due to a small test and training dataset.

This paper represents a description of Bulgarian verbal computer terms with a view to the specifics of their translation in English. The study employs a subset of 100 verbs extracted from the Bulgarian WordNet (BulNet) and from the internet. The analysis of their syntactic and semantic structure is a part of a study of the general lexis of Bulgarian. The aim of the paper is to (1) identify some problem areas of the description and translation of general lexis verbs, (2) offer an approach to the semantic description of metaphor-based terms from the perspective of Frame Semantics; (3) raise questions about the definition of general lexis with respect to Bulgarian and across languages.

Initial work has been done to address fake news detection and misrepresentation of news in the Bengali language. However, no work in Bengali yet addresses the identification of specific claims in social media news that falsely manipulate a related news article. At this point, this problem has been tackled in English and a few other languages, but not in the Bengali language. In this paper, we curate a dataset of social media content labeled with information manipulation relative to reference articles, called BanMANI. The dataset collection method we describe works around the limitations of the available NLP tools in Bangla. We expect these techniques will carry over to building similar datasets in other low-resource languages. BanMANI forms the basis both for evaluating the capabilities of existing NLP systems and for training or fine-tuning new models specifically on this task. In our analysis, we find that this task challenges current LLMs both under zero-shot and fine-tuned set- things

This study, submitted to the BUCC2023 shared task on bilingual term alignment in comparable specialised corpora, introduces a supervised, feature-based classification approach. The approach employs both static cross-lingual embeddings and contextual multilingual embeddings, combined with surface-level indicators such as Levenshtein distance and term length, as well as linguistic information. Results exhibit improved performance over previous methodologies, illustrating the merit of integrating diverse features. However, the error analysis also reveals remaining challenges.

We introduce calamanCy, an open-source toolkit for constructing natural language processing (NLP) pipelines for Tagalog. It is built on top of spaCy, enabling easy experimentation and integration with other frameworks. calamanCy addresses the development gap by providing a consistent API for building NLP applications and offering general-purpose multitask models with out-of-the-box support for dependency parsing, parts-of-speech (POS) tagging, and named entity recognition (NER). calamanCy aims to accelerate the progress of Tagalog NLP by consolidating disjointed resources in a unified framework.The calamanCy toolkit is available on GitHub: https://github.com/ljvmiranda921/calamanCy.

Jina Embeddings constitutes a set of high-performance sentence embedding models adept at translating textual inputs into numerical representations, capturing the semantics of the text. These models excel in applications like dense retrieval and semantic textual similarity. This paper details the development of Jina Embeddings, starting with the creation of high-quality pairwise and triplet datasets.It underlines the crucial role of data cleaning in dataset preparation, offers in-depth insights into the model training process, and concludes with a comprehensive performance evaluation using the Massive Text Embedding Benchmark (MTEB). Furthermore, to increase the model’s awareness of grammatical negation, we construct a novel training and evaluation dataset of negated and non-negated statements, which we make publicly available to the community.

Segmenting an address into meaningful components, also known as address parsing, is an essential step in many applications from record linkage to geocoding and package delivery. Consequently, a lot of work has been dedicated to develop accurate address parsing techniques, with machine learning and neural network methods leading the state-of-the-art scoreboard. However, most of the work on address parsing has been confined to academic endeavours with little availability of free and easy-to-use open-source solutions.This paper presents Deepparse, a Python open-source, extendable, fine-tunable address parsing solution under LGPL-3.0 licence to parse multinational addresses using state-of-the-art deep learning algorithms and evaluated on over 60 countries. It can parse addresses written in any language and use any address standard. The pre-trained model achieves average 99% parsing accuracies on the countries used for training with no pre-processing nor post-processing needed. Moreover, the library supports fine-tuning with new data to generate a custom address parser.

We present PyThaiNLP, a free and open-source natural language processing (NLP) library for Thai language implemented in Python. It provides a wide range of software, models, and datasets for Thai language. We first provide a brief historical context of tools for Thai language prior to the development of PyThaiNLP. We then outline the functionalities it provided as well as datasets and pre-trained language models. We later summarize its development milestones and discuss our experience during its development. We conclude by demonstrating how industrial and research communities utilize PyThaiNLP in their work. The library is freely available at https://github.com/pythainlp/pythainlp.

Knowledge extraction from scientific literature is a major issue, crucial to promoting transparency, reproducibility, and innovation in the research community. In this work, we present a novel approach towards the identification, extraction and analysis of dataset and code/software mentions within scientific literature. We introduce a comprehensive dataset, synthetically generated by ChatGPT and meticulously curated, augmented, and expanded with real snippets of scientific text from full-text publications in Computer Science using a human-in-the-loop process. The dataset contains snippets highlighting mentions of the two research artifact (RA) types: dataset and code/software, along with insightful metadata including their Name, Version, License, URL as well as the intended Usage and Provenance. We also fine-tune a simple Large Language Model (LLM) using Low-Rank Adaptation (LoRA) to transform the Research Artifact Analysis (RAA) into an instruction-based Question Answering (QA) task. Ultimately, we report the improvements in performance on the test set of our dataset when compared to other base LLM models. Our method provides a significant step towards facilitating accurate, effective, and efficient extraction of datasets and software from scientific papers, contributing to the challenges of reproducibility and reusability in scientific research.

Zelda Rose is a command line interface for pretraining transformer-based models. Its purpose is to enable an easy start for users interested in training these ubiquitous models, but unable or unwilling to engage with more comprehensive — but more complex — frameworks and the complex interactions between libraries for managing models, datasets and computations. Training a model requires no code on the user’s part and produce models directly compatible with the HuggingFace ecosystem, allowing quick and easy distribution and reuse. A particular care is given to lowering the cost of maintainability and future-proofing, by making the code as modular as possible and taking advantage of third-party libraries to limit ad-hoc code to the strict minimum.

Large language models (LLMs) have recently achieved human-level performance on a range of professional and academic benchmarks.The accessibility of these models has lagged behind their performance.State-of-the-art LLMs require costly infrastructure; are only accessible via rate-limited, geo-locked, and censored web interfaces; and lack publicly available code and technical reports.In this paper, we tell the story of GPT4All, a popular open source repository that aims to democratize access to LLMs.We outline the technical details of the original GPT4All model family, as well as the evolution of the GPT4All project from a single model into a fully fledged open source ecosystem.It is our hope that this paper acts as both a technical overview of the original GPT4All models as well as a case study on the subsequent growth of the GPT4All open source ecosystem.

Language model applications are becoming increasingly popular and complex, often including features like tool usage and retrieval augmentation. However, existing frameworks for such applications are often opinionated, deciding for developers how their prompts ought to be formatted and imposing limitations on customizability and reproducibility. To solve this we present Kani: a lightweight, flexible, and model-agnostic open-source framework for building language model applications. Kani helps developers implement a variety of complex features by supporting the core building blocks of chat interaction: model interfacing, chat management, and robust function calling. All Kani core functions are easily overridable and well documented to empower developers to customize functionality for their own needs. Kani thus serves as a useful tool for researchers, hobbyists, and industry professionals alike to accelerate their development while retaining interoperability and fine-grained control.

We present a case study describing our efforts to integrate the open source GECToR code and models into our production NLP pipeline that powers many of Educational Testing Service’s products and prototypes. The paper’s contributions includes a discussion of the issues we encountered during integration and our solutions, the overarching lessons we learned about integrating open source projects, and, last but not least, the open source contributions we made as part of the journey.

The ACL Anthology is a prime resource for research papers within computational linguistics and natural language processing, while continuing to be an open-source and community-driven project. Since Gildea et al. (2018) reported on its state and planned directions, the Anthology has seen major technical changes. We discuss what led to these changes and how they impact long-term maintainability and community engagement, describe which open-source data and software tools the Anthology currently provides, and provide a survey of literature that has used the Anthology as a main data source.

State-of-the-art language models like T5 have revolutionized the NLP landscape, but their computational demands hinder a large portion of the research community. To address this challenge, we present nanoT5, a specially-optimized PyTorch framework for efficient pre-training and fine-tuning of T5 models. Drawing on insights from optimizer differences and prioritizing efficiency, nanoT5 allows a T5-Base model to be pre-trained on a single GPU in just 16 hours, without any loss in performance. With the introduction of this open-source framework, we hope to widen the accessibility to language modelling research and cater to the community’s demand for more user-friendly T5 (Encoder-Decoder) implementations. We make our contributions, including configurations, codebase, pre-training insights, and pre-trained models, available to the public.

We present the AWARE-text package, an open-source software package for collecting textual data on Android mobile devices. This package allows for collecting short message service (SMS or text messages) and character-level keystrokes. In addition to collecting this raw data, AWARE-text is designed for on device lexicon processing, which allows one to collect standard textual-based measures (e.g., sentiment, emotions, and topics) without collecting the underlying raw textual data. This is especially important in the case of mobile phones, which can contain sensitive and identifying information. Thus, the AWARE-text package allows for privacy protection while simultaneously collecting textual information at multiple levels of granularity: person (lifetime history of SMS), conversation (both sides of SMS conversations and group chats), message (single SMS), and character (individual keystrokes entered across applications). Finally, the unique processing environment of mobile devices opens up several methodological and privacy issues, which we discuss.

Many machine translation toolkits make use of a data preparation step wherein raw data is transformed into a tensor format that can be used directly by the trainer. This preparation step is increasingly at odds with modern research and development practices because this process produces a static, unchangeable version of the training data, making common training-time needs difficult (e.g., subword sampling), time-consuming (preprocessing with large data can take days), expensive (e.g., disk space), and cumbersome (managing experiment combinatorics). We propose an alternative approach that separates the generation of data from the consumption of that data. In this approach, there is no separate pre-processing step; data generation produces an infinite stream of permutations of the raw training data, which the trainer tensorizes and batches as it is consumed. Additionally, this data stream can be manipulated by a set of user-definable operators that provide on-the-fly modifications, such as data normalization, augmentation or filtering. We release an open-source toolkit, SOTASTREAM, that implements this approach: https://github.com/marian-nmt/sotastream. We show that it cuts training time, adds flexibility, reduces experiment management complexity, and reduces disk space, all without affecting the accuracy of the trained models.

The paper discusses the Linguistic Field Data Management and Analysis System (LiFE), a new open-source, web-based software that systematises storage, management, annotation, analysis and sharing of linguistic data gathered from the field as well as that crawled from various sources on the web such as YouTube, Twitter, Facebook, Instagram, Blog, Newspaper, Wikipedia, etc. The app supports two broad workflows - (a) the field linguists’ workflow in which data is collected directly from the speakers in the field and analysed further to produce grammatical descriptions, lexicons, educational materials and possibly language technologies; (b) the computational linguists’ workflow in which data collected from the web using automated crawlers or digitised using manual or semi-automatic means, annotated for various tasks and then used for developing different kinds of language technologies. In addition to supporting these workflows, the app provides some additional features as well - (a) it allows multiple users to collaboratively work on the same project via its granular access control and sharing option; (b) it allows the data to be exported to various formats including CSV, TSV, JSON, XLSX, , PDF, Textgrid, RDF (different serialisation formats) etc as appropriate; (c) it allows data import from various formats viz. LIFT XML, XLSX, JSON, CSV, TSV, Textgrid, etc; (d) it allows users to start working in the app at any stage of their work by giving the option to either create a new project from scratch or derive a new project from an existing project in the app.The app is currently available for use and testing on our server (http://life.unreal-tece.co.in/) and its source code has been released under AGPL license on our GitHub repository (https://github.com/unrealtecellp/life). It is licensed under separate, specific conditions for commercial usage.

Rumour detection, particularly on social media, has gained popularity in recent years. The machine learning community has made significant contributions in investigating automatic methods to detect rumours on such platforms. However, these state-of-the-art (SoTA) models are often deployed by social media companies; ordinary end-users cannot leverage the solutions in the literature for their own rumour detection. To address this issue, we put forward a novel browser extension that allows these users to perform rumour detection on Twitter. Particularly, we leverage the performance from SoTA architectures, which has not been done previously. Initial results from a user study confirm that this browser extension provides benefit. Additionally, we examine the performance of our browser extension’s rumour detection model in a simulated deployment environment. Our results show that additional infrastructure for the browser extension is required to ensure its usability when deployed as a live service for Twitter users at scale.

Given the criticality and difficulty of reproducing machine learning experiments, there have been significant efforts in reducing the variance of these results. The ability to consistently reproduce results effectively strengthens the underlying hypothesis of the work and should be regarded as important as the novel aspect of the research itself. The contribution of this work is an open source framework that has the following characteristics: a) facilitates reproducing consistent results, b) allows hot-swapping features and embeddings without further processing and re-vectorizing the dataset, c) provides a means of easily creating, training and evaluating natural language processing deep learning models with little to no code changes, and d) is freely available to the community.

We describe the features of SeqScore, an MIT-licensed Python toolkit for working with named entity recognition (NER) data.While SeqScore began as a tool for NER scoring, it has been expanded to help with the full lifecycle of working with NER data: validating annotation, providing at-a-glance and detailed summaries of the data, modifying annotation to support experiments, scoring system output, and aiding with error analysis.SeqScore is released via PyPI (https://pypi.org/project/seqscore/) and development occurs on GitHub (https://github.com/bltlab/seqscore).

Reproducibility in scientific work has been becoming increasingly important in research communities such as machine learning, natural language processing, and computer vision communities due to the rapid development of the research domains supported by recent advances in deep learning. In this work, we present a significantly upgraded version of torchdistill, a modular-driven coding-free deep learning framework significantly upgraded from the initial release, which supports only image classification and object detection tasks for reproducible knowledge distillation experiments. To demonstrate that the upgraded framework can support more tasks with third-party libraries, we reproduce the GLUE benchmark results of BERT models using a script based on the upgraded torchdistill, harmonizing with various Hugging Face libraries. All the 27 fine-tuned BERT models and configurations to reproduce the results are published at Hugging Face, and the model weights have already been widely used in research communities. We also reimplement popular small-sized models and new knowledge distillation methods and perform additional experiments for computer vision tasks.

Captum is a comprehensive library for model explainability in PyTorch, offering a range of methods from the interpretability literature to enhance users’ understanding of PyTorch models. In this paper, we introduce new features in Captum that are specifically designed to analyze the behavior of generative language models. We provide an overview of the available functionalities and example applications of their potential for understanding learned associations within generative language models.

We present **nerblackbox**, a python library to facilitate the use of state-of-the-art transformer-based models for named entity recognition. It provides simple-to-use yet powerful methods to access data and models from a wide range of sources, for fully automated model training and evaluation as well as versatile model inference. While many technical challenges are solved and hidden from the user by default, **nerblackbox** also offers fine-grained control and a rich set of customizable features. It is thus targeted both at application-oriented developers as well as machine learning experts and researchers.

We present an open-source Python library for building and using datasets where inputs are clusters of textual data, and outputs are sequences of real values representing one or more timeseries signals. The news-signals library supports diverse data science and NLP problem settings related to the prediction of time series behaviour using textual data feeds. For example, in the news domain, inputs are document clusters corresponding to daily news articles about a particular entity, and targets are explicitly associated real-valued timeseries: the volume of news about a particular person or company, or the number of pageviews of specific Wikimedia pages. Despite many industry and research usecases for this class of problem settings, to the best of our knowledge, News Signals is the only open-source library designed specifically to facilitate data science and research settings with natural language inputs and timeseries targets. In addition to the core codebase for building and interacting with datasets, we also conduct a suite of experiments using several popular Machine Learning libraries, which are used to establish baselines for timeseries anomaly prediction using textual inputs.

Online data streams make training machine learning models hard because of distribution shift and new patterns emerging over time. For natural language processing (NLP) tasks that utilize a collection of features based on lexicons and rules, it is important to adapt these features to the changing data. To address this challenge we introduce PyTAIL, a python library, which allows a human in the loop approach to actively train NLP models. PyTAIL enhances generic active learning, which only suggests new instances to label by also suggesting new features like rules and lexicons to label. Furthermore, PyTAIL is flexible enough for users to accept, reject, or update rules and lexicons as the model is being trained. Finally, we simulate the performance of PyTAIL on existing social media benchmark datasets for text classification. We compare various active learning strategies on these benchmarks. The model closes the gap with as few as 10% of the training data. Finally, we also highlight the importance of tracking evaluation metric on remaining data (which is not yet merged with active learning) alongside the test dataset. This highlights the effectiveness of the model in accurately annotating the remaining dataset, which is especially suitable for batch processing of large unlabelled corpora. PyTAIL will be open sourced and available at https://github.com/socialmediaie/pytail.

One of the primary obstacles in the advancement of Natural Language Processing (NLP) technologies for low-resource languages is the lack of annotated datasets for training and testing machine learning models. In this paper, we present Antarlekhaka, a tool for manual annotation of a comprehensive set of tasks relevant to NLP. The tool is Unicode-compatible, language-agnostic, Web-deployable and supports distributed annotation by multiple simultaneous annotators. The system sports user-friendly interfaces for 8 categories of annotation tasks. These, in turn, enable the annotation of a considerably larger set of NLP tasks. The task categories include two linguistic tasks not handled by any other tool, namely, sentence boundary detection and deciding canonical word order, which are important tasks for text that is in the form of poetry. We propose the idea of sequential annotation based on small text units, where an annotator performs several tasks related to a single text unit before proceeding to the next unit. The research applications of the proposed mode of multi-task annotation are also discussed. Antarlekhaka outperforms other annotation tools in objective evaluation. It has been also used for two real-life annotation tasks on two different languages, namely, Sanskrit and Bengali. The tool is available at https://github.com/Antarlekhaka/code

The rise of ChatGPT1 has led to the development of artificial intelligence (AI) applications, particularly those that rely on large language models (LLMs). However, recalling LLM APIs can be expensive, and the response speed may slow down during LLMs’ peak times, causing frustration among developers. Potential solutions to this problem include using better LLM models or investing in more computing resources. However, these options may increase product development costs and decrease development speed. GPTCache2 is an open-source semantic cache that stores LLM responses to address this issue. When integrating an AI application with GPTCache, user queries are first sent to GPTCache for a response before being sent to LLMs like ChatGPT. If GPTCache has the answer to a query, it quickly returns the answer to the user without having to query the LLM. This approach saves costs on API recalls and makes response times much faster. For instance, integrating GPTCache with the GPT service offered by OpenAI can increase response speed 2-10 times when the cache is hit. Moreover, network fluctuations will not affect GPTCache’s response time, making it highly stable. This paper presents GPTCache and its architecture, how it functions and performs, and the use cases for which it is most advantageous.

Reinforcement learning from human feedback (RLHF) utilizes human feedback to better align large language models with human preferences via online optimization against a learned reward model. Current RLHF paradigms rely on Proximal Policy Optimization (PPO), which quickly becomes a challenge to implement and scale up to large architectures. To address this difficulty we created the trlX library as a feature-complete open-source framework for RLHF fine-tuning of models up to and exceeding 70 billion parameters. We implemented support for multiple types of distributed training including distributed data parallel, model sharded, as well as tensor, sequential, and pipeline parallelism.

Large language models (LLMs) have recently achieved human-level performance on a range of professional and academic benchmarks. Unfortunately, the explainability and accessibility of these models has lagged behind their performance. State-of-the-art LLMs require costly infrastructure, are only accessible via rate-limited, geo-locked, and censored web interfaces, and lack publicly available code and technical reports. Moreover, the lack of tooling for understanding the massive datasets used to train and produced by LLMs presents a critical challenge for explainability research. This talk will be an overview of Nomic AI’s efforts to address these challenges through its two core initiatives: GPT4All and Atlas

Natural Language Processing (NLP) tools have been rapidly adopted in political science for the study of conflict and violence. In this paper, we present an application to analyze various lethal and non-lethal events conducted by organized criminal groups and state forces in Mexico. Based on a large corpus of news articles in Spanish and a set of high-quality annotations, the application evaluates different Machine Learning (ML) algorithms and Large Language Models (LLMs) to classify documents and individual sentences, and to identify specific behaviors related to organized criminal violence and law enforcement efforts. Our experiments support the growing evidence that BERT-like models achieve outstanding classification performance for the study of organized crime. This application amplifies the capacity of conflict scholars to provide valuable information related to important security challenges in the developing world.

The method method presented in this paper uses a BERT model for classifying location mentions in event reporting news texts into two classes: a place of an event, called main location, or another location mention, called here secondary location. Our evaluation on articles, reporting protests, shows promising results and demonstrates the feasibility of our approach and the event geolocation task in general. We evaluate our method against a simple baseline and state of the art ML models and we achieve a significant improvement in all cases by using the BERT model. In contrast to other location classification approaches, we completelly avoid lingusitic pre processing and feature engineering, which is a pre-requisite for all multi-domain and multilingual applications.

Social media has become an established platform for people to organize and take offline actions, often in the form of civil unrest. Understanding these events can help support pro-democratic movements. The primary method to detect these events on Twitter relies on aggregating many tweets, but this includes many that are not relevant to the task. We propose a multi-instance learning (MIL) approach, which jointly identifies relevant tweets and detects civil unrest events. We demonstrate that MIL improves civil unrest detection over methods based on simple aggregation. Our best model achieves a 0.73 F1 on the Global Civil Unrest on Twitter (G-CUT) dataset.

Identifying cause-effect relations plays an integral role in the understanding and interpretation of natural languages. Furthermore, automated mining of causal relations from news and text about socio-political events is a stepping stone in gaining critical insights, including analyzing the scale, frequency and trends across timelines of events, as well as anticipating future ones. The Shared Task 3, part of the 6th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE @ RANLP 2023), involved the task of Event Causality Identification with Causal News Corpus. We describe our approach to Subtask 1, dealing with causal event classification, a supervised binary classification problem to annotate given event sentences with whether they contained any cause-effect relations. To help achieve this task, a BERT based architecture - RoBERTa was implemented. The results of this model are validated on the dataset provided by the organizers of this task.

Understanding causality is a core aspect of intelligence. The Event Causality Identification with Causal News Corpus Shared Task addresses two aspects of this challenge: Subtask 1 aims at detecting causal relationships in texts, and Subtask 2 requires identifying signal words and the spans that refer to the cause or effect, respectively. Our system, which is based on pre-trained transformers, stacked sequence tagging, and synthetic data augmentation, ranks third in Subtask 1 and wins Subtask 2 with an F1 score of 72.8, corresponding to a margin of 13 pp. to the second-best system.

Mapping ongoing news headlines to event-related classes in a rich knowledge base can be an important component in a knowledge-based event analysis and forecasting solution. In this paper, we present a methodology for creating a benchmark dataset of news headlines mapped to event classes in Wikidata, and resources for the evaluation of methods that perform the mapping. We use the dataset to study two classes of unsupervised methods for this task: 1) adaptations of classic entity linking methods, and 2) methods that treat the problem as a zero-shot text classification problem. For the first approach, we evaluate off-the-shelf entity linking systems. For the second approach, we explore a) pre-trained natural language inference (NLI) models, and b) pre-trained large generative language models. We present the results of our evaluation, lessons learned, and directions for future work. The dataset and scripts for evaluation are made publicly available.

Hate speech detection during times of war has become crucial in recent years, as evident with the recent Russo-Ukrainian war. In this paper, we present our submissions for both subtasks from the Multimodal Hate Speech Event Detec- tion contest at CASE 2023, RANLP 2023. We used pre-trained BERT models in both submis- sion, achieving a F1 score of 0.809 in subtask A, and F1 score of 0.567 in subtask B. In the first subtask, our result was not far from the first place, which led us to realize the lower impact of images in real-life memes about feel- ings, when compared with the impact of text. However, we observed a higher importance of images when targeting hateful feelings towards a specific entity. The source code to reproduce our results can be found at the github repository https://github.com/JesusASmx/OmeteotlAtCASE2023

Causal events play a crucial role in explaining the intricate relationships between the causes and effects of events. However, comprehending causal events within discourse, text, or speech poses significant semantic challenges. We propose a contrastive learning-based method in this submission to the Causal News Corpus - Event Causality Shared Task 2023, with a specific focus on SubTask1 centered on causal event classification. In our approach we pre-train our base model using Supervised Contrastive (SuperCon) learning. Subsequently, we fine-tune the pre-trained model for the specific task of causal event classification. Our experimentation demonstrates the effectiveness of our method, achieving a competitive performance, and securing the 2nd position on the leaderboard with an F1-Score of 84.36.

In this research paper, we propose a multimodal approach to hate speech detection, directed towards the identification of hate speech and its related targets. Our method uses logistic regression and support vector machines (SVMs) to analyse textual content extracted from social media platforms. We exploit natural language processing techniques to preprocess and extract relevant features from textual content, capturing linguistic patterns, sentiment, and contextual information.

Text-embedded images can serve as a means of spreading hate speech, propaganda, and extremist beliefs. Throughout the Russia-Ukraine war, both opposing factions heavily relied on text-embedded images as a vehicle for spreading propaganda and hate speech. Ensuring the effective detection of hate speech and propaganda is of utmost importance to mitigate the negative effect of hate speech dissemination. In this paper, we outline our methodologies for two subtasks of Multimodal Hate Speech Event Detection 2023. For the first subtask, hate speech detection, we utilize multimodal deep learning models boosted by ensemble learning and syntactical text attributes. For the second subtask, target detection, we employ multimodal deep learning models boosted by named entity features. Through experimentation, we demonstrate the superior performance of our models compared to all textual, visual, and text-visual baselines employed in multimodal hate speech detection. Furthermore, our models achieve the first place in both subtasks on the final leaderboard of the shared task.

Hate speech detection has emerged as a critical research area in recent years due to the rise of online social platforms and the proliferation of harmful content targeting individuals or specific groups.This task highlights the importance of detecting hate speech in text-embedded images.By leveraging deep learning models,this research aims to uncover the connection between hate speech and the entities it targets.

With a surge in the usage of social media postings to express opinions, emotions, and ideologies, there has been a significant shift towards the calibration of social media as a rapid medium of conveying viewpoints and outlooks over the globe. Concurrently, the emergence of a multitude of conflicts between two entities has given rise to a stream of social media content containing propaganda, hate speech, and inconsiderate views. Thus, the issue of monitoring social media postings is rising swiftly, attracting major attention from those willing to solve such problems. One such problem is Hate Speech detection. To mitigate this problem, we present our novel ensemble learning approach for detecting hate speech, by classifying text-embedded images into two labels, namely “Hate Speech” and “No Hate Speech” . We have incorporated state-of-art models including InceptionV3, BERT, and XLNet. Our proposed ensemble model yielded promising results with 75.21 and 74.96 as accuracy and F-1 score (respectively). We also present an empirical evaluation of the text-embedded images to elaborate on how well the model was able to predict and classify.

The paper presents a semantic model of protest events, called Semantic Interpretations of Protest Events (SemInPE). The analytical framework used for building the semantic representations is inspired by the object-oriented paradigm in computer science and a cognitive approach to the linguistic analysis. The model is a practical application of the Unified Eventity Representation (UER) formalism, which is based on the Unified Modeling Language (UML). The multi-layered architecture of the model provides flexible means for building the semantic representations of the language objects along a scale of generality and specificity. Thus, it is a suitable environment for creating the elements of ontologies on various topics and for different languages.

The emergence of social media and e-commerce platforms enabled the perpetrator to spread negativity and abuse individuals or organisations worldwide rapidly. It is critical to detect hate speech in both visual and textual content so that it may be moderated or excluded from online platforms to keep it sound and safe for users. However, multimodal hate speech detection is a complex and challenging task as people sarcastically present hate speech and different modalities i.e., image and text are involved in their content. This paper describes our participation in the CASE 2023 multimodal hate speech event detection task. In this task, the objective is to automatically detect hate speech and its target from the given text-embedded image. We proposed a transformer-based multimodal hierarchical fusion model to detect hate speech present in the visual content. We jointly fine-tune a language and a vision pre-trained transformer models to extract the visual-contextualized features representation of the text-embedded image. We concatenate these features and fed them to the multi-sample dropout strategy. Moreover, the contextual feature vector is fed into the BiLSTM module and the output of the BiLSTM module also passes into the multi-sample dropout. We employed arithmetic mean fusion to fuse all sample dropout outputs that predict the final label of our proposed method. Experimental results demonstrate that our model obtains competitive performance and ranked 5th among the participants

Cause-effect relationships play a crucial role in human cognition, and distilling cause-effect relations from text helps in ameliorating causal networks for predictive tasks. There are many NLP applications that can benefit from this task, including natural language-based financial forecasting, text summarization, and question-answering. However, due to the lack of syntactic clues, the ambivalent semantic meaning of words, complex sentence structure, and implicit meaning of numerical entities in the text make it one of the challenging tasks in NLP. To address these challenges, CASE-2023 introduced a shared task 3 task focusing on event causality identification with causal news corpus. In this paper, we demonstrate our participant systems for this task. We leverage two transformers models including DeBERTa and Twitter-RoBERTa along with the weighted average fusion technique to tackle the challenges of subtask 1 where we need to identify whether a text belongs to either causal or not. For subtask 2 where we need to identify the cause, effect, and signal tokens from the text, we proposed a unified neural network of DeBERTa and DistilRoBERTa transformer variants with contrastive learning techniques. The experimental results showed that our proposed method achieved competitive performance among the participants’ systems.

Event extraction from textual data is a NLP research task relevant to a plethora of domains. Most approaches aim to recognize events from a predefined event schema, consisting of event types and their corresponding arguments. For domains, such as disinformation, where new event types emerge frequently, there is a need to adapt such fixed event schemas to accommodate for new event types. We present NEXT (New Event eXTraction) - a resource-sparse approach to extending a close-domain model to novel event types, that requires a very small number of annotated samples for fine-tuning performed on a single GPU. Furthermore, our results suggest that this approach is suitable not only for extraction of new event types, but also for recognition of existing event types, as the use of this approach on a new dataset leads to improved recall for all existing events while retaining precision.

The scarcity of data poses a significant challenge in closed-domain event extraction, as is common in complex NLP tasks. This limitation primarily arises from the intricate nature of the annotation process. To address this issue, we present a multi-task model structure and training approach that leverages the additional data, which is found as not having any event information at document and sentence levels, generated during the event annotation process. By incorporating this supplementary data, our proposed framework demonstrates enhanced robustness and, in some scenarios, improved performance. A particularly noteworthy observation is that including only negative documents in addition to the original data contributes to performance enhancement. Our findings offer promising insights into leveraging extra data to mitigate data scarcity challenges in closed-domain event extraction.

Hate speech has emerged as a pressing issue on social media platforms, fueled by the increasing availability of multimodal data and easy internet access. Addressing this problem requires collaborative efforts from researchers, policymakers, and online platforms. In this study, we investigate the detection of hate speech in multimodal data, comprising text-embedded images, by employing advanced deep learning models. The main objective is to identify effective strategies for hate speech detection and content moderation. We conducted experiments using four state-of-the-art classifiers: XLM-Roberta-base, BiLSTM, XLNet base cased, and ALBERT, on the CrisisHateMM[4] dataset, consisting of over 4700 text-embedded images related to the Russia-Ukraine conflict. The best findings reveal that XLM-Roberta-base exhibits superior performance, outperforming other classifiers across all evaluation metrics, including an impressive F1 score of 84.62 for sub-task 1 and 69.73 for sub-task 2. The future scope of this study lies in exploring multimodal approaches to enhance hate speech detection accuracy, integrating ethical considerations to address potential biases, promoting fairness, and safeguarding user rights. Additionally, leveraging larger and more diverse datasets will contribute to developing more robust and generalised hate speech detection solutions.

The Event Causality Identification Shared Task of CASE 2023 is the second iteration of a shared task centered around the Causal News Corpus. Two subtasks were involved: In Subtask 1, participants were challenged to predict if a sentence contains a causal relation or not. In Subtask 2, participants were challenged to identify the Cause, Effect, and Signal spans given an input causal sentence. For both subtasks, participants uploaded their predictions for a held-out test set, and ranking was done based on binary F1 and macro F1 scores for Subtask 1 and 2, respectively. This paper includes an overview of the work of the ten teams that submitted their results to our competition and the six system description papers that were received. The highest F1 scores achieved for Subtask 1 and 2 were 84.66% and 72.79%, respectively.

Ensuring the moderation of hate speech and its targets emerges as a critical imperative within contemporary digital discourse. To facilitate this imperative, the shared task Multimodal Hate Speech Event Detection was organized in the sixth CASE workshop co-located at RANLP 2023. The shared task has two subtasks. The sub-task A required participants to pose hate speech detection as a binary problem i.e. they had to detect if the given text-embedded image had hate or not. Similarly, sub-task B required participants to identify the targets of the hate speech namely individual, community, and organization targets in text-embedded images. For both sub-tasks, the participants were ranked on the basis of the F1-score. The best F1-score in sub-task A and sub-task B were 85.65 and 76.34 respectively. This paper provides a comprehensive overview of the performance of 13 teams that submitted the results in Subtask A and 10 teams in Subtask B.

The purpose of the shared task 2 at the Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE) 2023 workshop was to test the abilities of the participating models and systems to detect and geocode armed conflicts events in social media messages from Telegram channels reporting on the Russo Ukrainian war. The evaluation followed an approach which was introduced in CASE 2021 (Giorgi et al., 2021): For each system we consider the correlation of the spatio-temporal distribution of its detected events and the events identified for the same period in the ACLED (Armed Conflict Location and Event Data Project) database (Raleigh et al., 2010). We use ACLED for the ground truth, since it is a well established standard in the field of event extraction and political trend analysis, which relies on human annotators for the encoding of security events using a fine grained taxonomy. Two systems participated in this shared task, we report in this paper on both the shared task and the participating systems.

We provide a summary of the sixth edition of the CASE workshop that is held in the scope of RANLP 2023. The workshop consists of regular papers, three keynotes, working papers of shared task participants, and shared task overview papers. This workshop series has been bringing together all aspects of event information collection across technical and social science fields. In addition to contributing to the progress in text based event extraction, the workshop provides a space for the organization of a multimodal event information collection task.

This paper investigates whether current large language models exhibit biases in logical reasoning, similar to humans. Specifically, we focus on syllogistic reasoning, a well-studied form of inference in the cognitive science of human deduction. To facilitate our analysis, we introduce a dataset called NeuBAROCO, originally designed for psychological experiments that assess human logical abilities in syllogistic reasoning. The dataset consists of syllogistic inferences in both English and Japanese. We examine three types of biases observed in human syllogistic reasoning: belief biases, conversion errors, and atmosphere effects. Our findings demonstrate that current large language models struggle more with problems involving these three types of biases.

While many natural language inference (NLI) datasets target certain semantic phenomena, e.g., negation, tense & aspect, monotonicity, and presupposition, to the best of our knowledge, there is no NLI dataset that involves diverse types of spatial expressions and reasoning. We fill this gap by semi-automatically creating an NLI dataset for spatial reasoning, called SpaceNLI. The data samples are automatically generated from a curated set of reasoning patterns (see Figure 1), where the patterns are annotated with inference labels by experts. We test several SOTA NLI systems on SpaceNLI to gauge the complexity of the dataset and the system’s capacity for spatial reasoning. Moreover, we introduce a Pattern Accuracy and argue that it is a more reliable and stricter measure than the accuracy for evaluating a system’s performance on pattern-based generated data samples. Based on the evaluation results we find that the systems obtain moderate results on the spatial NLI problems but lack consistency per inference pattern. The results also reveal that non-projective spatial inferences (especially due to the “between” preposition) are the most challenging ones.

Recent advances in large language models (LLMs) and LLM-driven chatbots, such as ChatGPT, have sparked interest in the extent to which these artificial systems possess human-like linguistic abilities. In this study, we assessed ChatGPT’s pragmatic capabilities by conducting three preregistered experiments focused on its ability to compute pragmatic implicatures. The first experiment tested whether ChatGPT inhibits the computation of generalized conversational implicatures (GCIs) when explicitly required to process the text’s truth-conditional meaning. The second and third experiments examined whether the communicative context affects ChatGPT’s ability to compute scalar implicatures (SIs). Our results showed that ChatGPT did not demonstrate human-like flexibility in switching between pragmatic and semantic processing. Additionally, ChatGPT’s judgments did not exhibit the well-established effect of communicative context on SI rates.

The two contrasting approaches are end-to-end neural NLI systems and linguistically-oriented NLI pipelines consisting of modules such as neural CCG parsers and theorem provers. The latter, however, faces the challenge of integrating the neural models used in the syntactic and semantic components. RNNGs are frameworks that can potentially fill this gap, but conventional RNNGs adopt CFG as the syntactic theory. To address this issue, we implemented RNN-CCG, a syntactic parser that replaces CFG with CCG. We then conducted experiments comparing RNN-CCG to RNNGs with/without POS tags and evaluated their behavior as a first step towards building an NLI system based on RNN-CCG.

This paper considers how the kind of formal semantic objects used in TTR (a theory of types with records, Cooper 2013) might be related to the vector representations used in Eliasmith (2013). An advantage of doing this is that it would immediately give us a neural representation for TTR objects as Eliasmith relates vectors to neural activity in his semantic pointer architecture (SPA). This would be an alternative using convolution to the suggestions made by Cooper (2019) based on the phasing of neural activity. The project seems potentially hopeful since all complex TTR objects are constructed from labelled sets (essentially sets of ordered pairs consisting of labels and values) which might be seen as corresponding to the representation of structured objects which Eliasmith achieves using superposition and circular convolution.

Triadic representations that temporally order events and states are described, consisting of strings and sets of strings of bounded but refinable granularities. The strings are compressed according to J.A. Wheeler’s dictum it-from-bit, with bits given by statives and non-statives alike. A choice of vocabulary and constraints expressed in that vocabulary shape representations of cause-and-effect with deformations characteristic, Mumford posits, of patterns at various levels of cognitive processing. These deformations point to an ongoing process of learning, formulated as grammatical inference of finite automata, structured around Goguen and Burstall’s institutions.

Previous work has predominantly focused on monolingual English semantic parsing. We, instead, explore the feasibility of Chinese semantic parsing in the absence of labeled data for Chinese meaning representations. We describe the pipeline of automatically collecting the linearized Chinese meaning representation data for sequential-to-sequential neural networks. We further propose a test suite designed explicitly for Chinese semantic parsing, which provides fine-grained evaluation for parsing performance, where we aim to study Chinese parsing difficulties. Our experimental results show that the difficulty of Chinese semantic parsing is mainly caused by adverbs. Realizing Chinese parsing through machine translation and an English parser yields slightly lower performance than training a model directly on Chinese data.

Indigenous language expertise is not encoded in written text in the same way as it is for languages that have a long literal tradition. In many cases it is, on the contrary, mostly conserved orally. Therefore the evaluation of neural MT systems solely based on an algorithm learning from written texts is not adequate to measure the quality of a system that is used by the language community. If extensively using tools based on a big amount of non-native language this can even contribute to language change in a way that is not desired by the language community. It can also pollute the internet with automatically created texts that outweigh native texts. We propose a manual evaluation method focusing on flow and content separately, and additionally we use existing rule-based NLP to evaluate other factors such as spelling, grammar and grammatical richness. Our main conclusion is that language expertise of a native speaker is necessary to properly evaluate a given system. We test the method by manually evaluating two neural MT tools for an indigenous low resource language. We present an experiment on two different neural translations to and from North Sámi, an indigenous language of North Europe.

Human evaluation plays a crucial role in Natural Language Processing (NLP) as it assesses the quality and relevance of developed systems, thereby facilitating their enhancement. However, the absence of widely accepted human evaluation metrics in NLP hampers fair comparisons among different systems and the establishment of universal assessment standards. Through an extensive analysis of existing literature on human evaluation metrics, we identified several gaps in NLP evaluation methodologies. These gaps served as motivation for developing our own hierarchical evaluation framework. The proposed framework offers notable advantages, particularly in providing a more comprehensive representation of the NLP system’s performance. We applied this framework to evaluate the developed Machine Reading Comprehension system, which was utilized within a human-AI symbiosis model. The results highlighted the associations between the quality of inputs and outputs, underscoring the necessity to evaluate both components rather than solely focusing on outputs. In future work, we will investigate the potential time-saving benefits of our proposed framework for evaluators assessing NLP systems.

In both the translation industry and translation education, analytic and systematic assessment of translations plays a vital role. However, due to lack of a scheme for describing differences between translations, such assessment has been realized only in an ad-hoc manner. There is prior work on a scheme for describing differences between translations, but it has coverage and objectivity issues. To alleviate these issues and realize more fine-grained analyses, we developed an improved scheme by referring to diverse types of translations and adopting hierarchical linguistic units for analysis, taking English-to-Japanese translation as an example.

This paper presents an overview of, and the results from, the 2023 Shared Task on Reproducibility of Evaluations in NLP (ReproNLP’23), following on from two previous shared tasks on reproducibility of evaluations in NLG, ReproGen’21 and ReproGen’22. This shared task series forms part of an ongoing research programme designed to develop theory and practice of reproducibility assessment in NLP and machine learning, all against a background of an interest in reproducibility that con- tinues to grow in the two fields. This paper describes the ReproNLP’23 shared task, summarises results from the reproduction studies submitted, and provides comparative analysis of the results.

This paper presents a human evaluation reproduction study regarding the data-to-text generation task. The evaluation focuses in counting the supported and contradicting facts generated by a neural data-to-text model with a macro planning stage. The model is tested generating sport summaries for the ROTOWIRE dataset. We first describe the approach to reproduction that is agreed in the context of the ReproHum project. Then, we detail the entire configuration of the original human evaluation and the adaptations that had to be made to reproduce such an evaluation. Finally, we compare the reproduction results with those reported in the paper that was taken as reference.

Human evaluation is crucial for NLG systems as it provides a reliable assessment of the quality, effectiveness, and utility of generated language outputs. However, concerns about the reproducibility of such evaluations have emerged, casting doubt on the reliability and generalisability of reported results. In this paper, we present the findings of a reproducibility study on a data-to-text system, conducted under two conditions: (1) replicating the original setup as closely as possible with evaluators from AMT, and (2) replicating the original human evaluation but this time, utilising evaluators with a background in academia. Our experiments show that there is a loss of statistical significance between the original and reproduction studies, i.e. the human evaluation results are not reproducible. In addition, we found that employing local participants led to more robust results. We finally discuss lessons learned, addressing the challenges and best practices for ensuring reproducibility in NLG human evaluations.

This paper is part of the larger ReproHum project, where different teams of researchers aim to reproduce published experiments from the NLP literature. Specifically, ReproHum focuses on the reproducibility of human evaluation studies, where participants indicate the quality of different outputs of Natural Language Generation (NLG) systems. This is necessary because without reproduction studies, we do not know how reliable earlier results are. This paper aims to reproduce the second human evaluation study of Puduppully & Lapata (2021), while another lab is attempting to do the same. This experiment uses best-worst scaling to determine the relative performance of different NLG systems. We found that the worst performing system in the original study is now in fact the best performing system across the board. This means that we cannot fully reproduce the original results. We also carry out alternative analyses of the data, and discuss how our results may be combined with the other reproduction study that is carried out in parallel with this paper.

This paper presents a partial reproduction study of Data-to-text Generation with Macro Planning by Puduppully et al. (2021). This work was conducted as part of the ReproHum project, a multi-lab effort to reproduce the results of NLP papers incorporating human evaluations. We follow the same instructions provided by the authors and the ReproHum team to the best of our abilities. We collect preference ratings for the following evaluation criteria in order: conciseness, coherence, and grammaticality. Our results are highly correlated with the original experiment. Nonetheless, we believe the presented results are insufficent to conclude that the Macro system proposed and developed by the original paper is superior compared to other systems. We suspect combining our results with the three other reproductions of this paper through the ReproHum project will paint a clearer picture. Overall, we hope that our work is a step towards a more transparent and reproducible research landscape.

There is a growing concern regarding the reproducibility of human evaluation studies in NLP. As part of the ReproHum campaign, we conducted a study to assess the reproducibility of a recent human evaluation study in NLP. Specifically, we attempted to reproduce a human evaluation of a novel approach to enhance Role-Oriented Dialogue Summarization by considering the influence of role interactions. Despite our best efforts to adhere to the reported setup, we were unable to reproduce the statistical results as presented in the original paper. While no contradictory evidence was found, our study raises questions about the validity of the reported statistical significance results, and/or the comprehensiveness with which the original study was reported. In this paper, we provide a comprehensive account of our reproduction study, detailing the methodologies employed, data collection, and analysis procedures. We discuss the implications of our findings for the broader issue of reproducibility in NLP research. Our findings serve as a cautionary reminder of the challenges in conducting reproducible human evaluations and prompt further discussions within the NLP community.

This paper reports a reproduction study of the human evaluation of role-oriented dialogue summarization models, as part of the ReproNLP Shared Task 2023 on Reproducibility of Evaluations in NLP. We outline the disparities between the original study’s experimental design and our reproduction study, along with the outcomes obtained. The inter-annotator agreement within the reproduction study is observed to be lower, measuring 0.40 as compared to the original study’s 0.48. Among the six conclusions drawn in the original study, four are validated in our reproduction study. We confirm the effectiveness of the proposed approach on the overall metric, albeit with slightly poorer relative performance compared to the original study. Furthermore, we raise an open-ended inquiry: how can subjective practices in the original study be identified and addressed when conducting reproduction studies?

How reliable are human evaluation results? Is it possible to replicate human evaluation? This work takes a closer look at the evaluation of the output of a Text-to-Speech (TTS) system. Unfortunately, our results indicate that human evaluation is not as straightforward to replicate as expected. Additionally, we also present results on reproducing the technical background of the TTS system and discuss potential reasons for the reproduction failure.

This paper describes the reproduction of a human evaluation in Language-Agnostic Meta- Learning for Low-Resource Text-to-Speech with Articulatory Features reported in Lux and Vu (2022). It is a contribution to the ReproNLP 2023 Shared Task on Reproducibility of Evaluations in NLP. The original evaluation assessed the naturalness of audio generated by different Text-to-Speech (TTS) systems for German, and our goal was to repeat the experiment with a different set of evaluators. We reproduced the evaluation based on data and instructions provided by the original authors, with some uncertainty concerning the randomisation of question order. Evaluators were recruited via email to relevant mailing lists and we received 157 responses over the course of three weeks. Our initial results show low reproducibility, but when we assume that the systems of the original and repeat evaluation experiment have been transposed, the reproducibility assessment improves markedly. We do not know if and at what point such a transposition happened; however, an initial analysis of our audio and video files provides some evidence that the system assignment in our repeat experiment is correct.

This work presents our efforts to reproduce the results of the human evaluation experiment presented in the paper of Vamvas and Sennrich (2022), which evaluated an automatic system detecting over- and undertranslations (translations containing more or less information than the original) in machine translation (MT) outputs. Despite the high quality of the documentation and code provided by the authors, we discuss some problems we found in reproducing the exact experimental setup and offer recommendations for improving reproducibility. Our replicated results generally confirm the conclusions of the original study, but in some cases statistically significant differences were observed, suggesting a high variability of human annotation.

This report describes a reproduction of a human evaluation study evaluating automatically detected over- and undertranslations obtained using neural machine translation approaches. While the scope of the original study is much broader, a human evaluation is included as part of its system evaluation. We attempt an exact reproduction of this human evaluation, pertaining to translations on the the English-German language pair. While encountering minor logistical challenges, with all the source material being publicly available and some additional instructions provided by the original authors, we were able to reproduce the original experiment with only minor differences in the results.

We reproduced the human-based evaluation of the continuation of narratives task presented by Chakrabarty et al. (2022). This experiment is performed as part of the ReproNLP Shared Task on Reproducibility of Evaluations in NLP (Track C). Our main goal is to reproduce the original study under conditions as similar as possible. Specifically, we follow the original experimental design and perform human evaluations of the data from the original study, while describing the differences between the two studies. We then present the results of these two studies together with an analysis of similarities between them. Inter-annotator agreement (Krippendorff’s alpha) in the reproduction study is lower than in the original study, while the human evaluation results of both studies have the same trends, that is, our results support the findings in the original study.

We describe in this paper an attempt to reproduce some of the human of evaluation results from the paper “It’s not Rocket Science: Interpreting Figurative Language in Narratives”. In particular, we describe the methodology used to reproduce the chosen human evaluation, the challenges faced, and the results that were gathered. We will also make some recommendations on the learnings obtained from this reproduction attempt and what improvements are needed to enable more robust reproductions of future NLP human evaluations.

Position paper for YRRSDS 2023

In this position paper, I will present the research interests in my PostDoc on safety and robustness specific to conversational AI, including then relevant overlap from my PhD.

Speech production is nuanced and unique to every individual, but today’s Spoken Dialogue Systems (SDSs) are trained to use general speech patterns to successfully improve performance on various evaluation metrics. However, these patterns do not apply to certain user groups - often the very people that can benefit the most from SDSs. For example, people with dementia produce more disfluent speech than the general population. The healthcare domain is now a popular setting for spoken dialogue and human-robot interaction research. This trend is similar when observing company behaviour. Charities promote industry voice assistants, the creators are getting HIPAA compliance, and their features sometimes target vulnerable user groups. It is therefore critical to adapt SDSs to be more accessible.

This research encompasses a comprehensive exploration of Spoken Dialogue Systems (SDSs) in the manufacturing sector. It begins by establishing a conceptual architecture and taxonomy to guide the design and selection of SDS elements. Real case applications, including worker safety and cybersecurity support, validate the research findings and highlight areas for improvement. Looking ahead, the study delves into the potential of Large Language Models (LLMs) and multi-modal applications. Emphasizing the importance of extreme personalization, the study highlights the need to cater to the diverse qualifications and preferences of workers. Additionally, it investigates the integration of SDSs with other sensory modalities, such as images, videos, and augmented or virtual reality scenarios, to enhance the user experience and productivity. The research also addresses crucial considerations related to knowledge base optimization. It examines semantic variations of words across different application contexts, the continuous updating of procedures and data, and the adaptability of SDSs to diverse dialects and linguistic abilities, particularly in low-schooling personnel scenarios. Privacy, industrial protection, and ethical concerns in the era of LLMs and external players like OpenAI are given due attention. The study explores the boundaries of knowledge that conversational systems should possess, advocating for transparency, explainability, and responsible data handling practices.

The process of “conversational grounding” is an interactive process that has been studied extensively in cognitive science, whereby participants in a conversation check to make sure their interlocutors understand what is being referred to. This interactive process uses multiple modes of communication to establish the information between the participants. This could include information provided through eye-gaze, head movements, intonation in speech, along with the content of the speech. While the process is essential to successful communication between humans and between humans and machines, work needs to be done on testing and building the capabilities of the current dialogue system in managing conversational grounding, especially in multimodal medium of communication. Recent work such as Benotti and Blackburn have shown the importance of conversational grounding in dialog systems and how current systems fail in them. This is essential for the advancement of Embodied Conversational Agents and Social Robots. Thus my PhD project aims to test, understand and improve the functioning of current dialog models with respect to Conversational Grounding.

My research interests focus on natural language generation (NLG) regarding how to make system outputs more intuitive and comprehensible for the human-user and conversational entrainment and alignment from the perspective of how dialogue systems could or should personalize its responses to the human user. As it relates to NLG, my current work focuses on training a system to auto-generate comments for SQL queries produced by a Text-to-SQL parser. The goal is to make the connection between technical SQL language and the user’s question more transparent. My linguistic training lies primarily at the intersection of computational and socio-linguistics. As such, my curiosities in conversational entrainment and alignment focus on the extent to which conversational agents can or should adjust their language based on human characteristics such as age, race, or gender.

Position paper for YRRSDS 2023

Large language models (LLMs) have brought about a significant transformation in spoken dialogue systems (SDSs). It is anticipated that these systems will be implemented into diverse robotic applications and employed in a variety of social settings. The author presents research interest with the aim of realizing social SDSs from multiple perspectives, including task design, turn-taking mechanisms, and evaluation methodologies. Additionally, future research in social SDSs should delve into a deeper understanding of user mental states and a relationship with society via multi-party conversations. Finally, the author suggests topics for discussion regarding the future directions of SDS researchers in the LLM era.

Many companies use dialogue systems for their customer service, and although there has been a rise in the usage of these systems (Costello and LoDolce, 2022), many of these systems still face challenges in comprehending and properly responding to the customer (Følstadet al., 2021). In our project we aim to figure out how to develop and improve these conversational agents. Part of this project (detailed in this paper) will focus on the detection of breakdown patterns and the possible solutions (repairs) to mitigate negative results of these errors.

Position paper on the intersection between chitchat and task-oriented dialogues (TODs), with a focus on integrating capabilities typically associated with chitchat systems into task-oriented agents.

This is my position paper for YRRSDS 2023. In it, I write about the details of my research interests as well as past, current and future projects, talk about the status of spoken dialogue system research, include a short bio, and suggest topics for discussion.

This position paper is an overview of author’s main research interests and work considering deep learning techniques in audio classification, sign languages, and multimodality in dialogue systems. Author also shares her opinion on current and future research considering dialogue agents, and suggests topics for discussion panels.

Task-Oriented Dialogue (TOD) systems provide interactive assistance to a user in order to accomplish a specific task such as making a reservation at a restaurant or booking a room in a hotel. Speech presents itself as a natural interface for TOD systems. A typical approach to implement them is to use a modular architecture (Gao et al., 2018). A core component of such dialogue systems is Spoken Language Understanding (SLU) whose goal is to extract the relevant information from the user’s utterances. While spoken dialogue was the focus of earlier work (Williams et al., 2013; Henderson et al., 2014), recent work has focused on text inputs with no regard for the specificities of spoken language (Wu et al., 2019; Heck et al., 2020; Feng et al., 2021). However, this approach fails to account for the differences between written and spoken language (Faruqui and Hakkani-Tür, 2022) such as disfluencies. My research focuses on Spoken Language Understanding in the context of Task-Oriented Dialogue. More specifically I am interested in the two following research directions: • Annotation schema for spoken TODs, • Integration of dialogue history for contextually coherent predictions.

My position paper for the YRRSDS 2023 workshop.

My PhD focuses on conversational agents for behaviour change, with a focus on the feasibility of applying Large Language Models (LLMs) such as GPT-4 in this context.

My primary research focus lies in the domain of Text Style Transfer (TST), a fascinating area within Natural Language Processing (NLP). TST involves the transfor- mation of text into a desired style while approximately preserving its underlying content. In my research, I am also driven by the goal of incorporating TST techniques into NLP systems, particularly within the realm of dia- logue systems. I am intrigued by the concept of Stylized Dialog Response Generation, which aims to enhance the versatility and adaptability of dialog systems in generat- ing text responses with specific style attributes. By ad- vancing our understanding of TST and its integration into dialogue systems, my research seeks to contribute to the broader field of human-computer interaction. Through the development of robust and versatile dialogue systems with enhanced style transfer capabilities, we can facili- tate more engaging and personalized conversational experiences.

A brief introduction to author’s keyinterests and research topics which are: multimodal dialogue systems and impact of data augmentation to NLU performance. In addition to that the author shares his biography and view on the future of dialogue assistants.

This submission discusses my research interests in two areas: measuring user satisfaction in goal-oriented dialogue systems and exploring the potential of multi-modal interactions. For goal-oriented dialogue systems, I focus on evaluating and enhancing user satisfaction throughout the interaction process, aiming to propose innovative strategies and address the limitations of existing evaluation techniques. Additionally, I explore the benefits of multi-modal dialogue systems, highlighting their ability to provide more natural and immersive conversations by incorporating various communication modes such as speech, text, gestures, and visuals.

My research interests broadly lie in the area of Information Extraction from Spoken Dialogue, with a spacial focus on state modeling, anaphora resolution, program synthesis & planning, and intent classification in goal-oriented conversations. My aim is to create embedded dialogue systems that can interact with humans in a collaborative setup to solve tasks in a digital/non-digital environment. Most of the goal-oriented conversations usually involve experts and a laypersons. The aim for the expert is to consider all the information provided by the layperson, identify the underlying set of issues or intents, and prescribe solutions. While human experts are very good at extracting such information, AI agents (that build up most of the automatic dialog systems today) not so much. Most of the existing assistants (or chatbots) only consider individual utterances and do not ground them in the context of the dialogue. My work in this direction has focused on making these systems more effective at extracting the most relevant information from the dialogue to help the human user reach their end-goal.

My research interests lie in the area of modelling natural and human-like conversations, with a special focus on emotions in task-oriented dialogue (ToD) systems. ToD systems need to produce semantically and grammatically correct responses to fulfil the user’s goal. Being able to perceive and express emotions pushes them one more step towards achieving human-likeness. To begin with, I constructed a dataset with meaningful emotion labels as well as a wide coverage of emotions and linguistic features in ToDs. Then, I improved emotion recognition in conversations (ERC) in the task-oriented domain by exploiting key characteristics of ToDs. Currently, I am working towards enhancing ToD systems with emotions.

I am broadly interested in evaluation of dialogue systems, in all its many facets: The data they are trained on, their ability to perform a task successfully, their skills with respect to various dialogue phenomena, their resemblance to human cognitive processes, and their ethical and societal impact. More specifically, my research topics focus on understanding the possibilities and limits of current multimodal neural network-based models to incrementally encode information for natural language understanding in general and also for building common ground and asking for clarification. Besides, I am interested in dialogue games as a means to elicit and collect dialogue data and to evaluate the abilities of dialogue models.

My research work centers on how to enable a human-like interaction through generating contextual, emotional or proactive responses, both in task-oriented and in chitchat spoken dialogue systems (SDSs), because natural lan- guage generation (NLG) is an indispensable component in SDSs and can directly affect the user interactive expe- rience of the entire dialogue system. In addition to NLG, I am also interested in natural language understanding (NLU), as it plays a crucial role in SDSs and is a prerequisite for dialogue systems to generate replies.

The author’s objective centers around developing a spoken dialogue system (SDS) that can emulate the cognitive and conversational qualities of a human friend. Key attributes such as empathy, knowledge/causality reasoning, and personality are integral components of human interaction. The proposed approach involves the creation of an Empathy-enriched SDS, capable of comprehending human emotions and circumstances, thus providing companionship and assistance akin to a trusted friend. Additionally, the Causality-reasoning for SDS aims to ground the system in commonsense knowledge and equip it with the ability to reason about causalities, such as predicting user desires/reactions and system intentions/reactions, thereby enhancing the system’s intelligence and human-like behavior. Finally, the concept of a Personality-conditioned SDS involves enabling systems to exhibit distinct personalities, further enhancing the naturalness of human-robot interaction.

This position paper describes my research interests, spoken dialogue system research, and suggested topics for discussion.

With the rise of the metaverse, immersive multimodal conversation has attracted more and more researchers’ attention. Multimodal contexts will become more important for human-computer interaction in the metaverse, especially in shopping domain. Unlike traditional conversation tasks, immersive multimodal conversation has challenges such as multimodal ambiguous candidate identification and multimodal coreference resolution, which makes it more difficult to dialog state tracking and response generation, as described in SIMMC 2.1 challenge, a part of DSTC11. In particular, as the number of objects in the scene increases, the difficulty will increase dramatically. We proposed a prompt-based multi-task learning Encoder-Decoder, in which different subtasks use different prompts to make the model tend to focus on the current subtask. We achieve the winner in ambiguous candidates indentification and runner-up in multimodal coreference resolution (MM-Coref), multimodal dialog state tracking (MM-DST) and assistant response generation. Our code and model are made publicly available at https://github.com/scutcyr/dstc11-simmc2.1-scut-bds-lab.

Recently, research using multimodal datasets containing image and text information has been conducted actively. One of them is the SIMMC2.1 dataset. It is a more complicated dataset than answering a conversation using only text because it should predict an answer after understanding the relationship between images and text. Therefore, there are limitations to answering a conversation only using text-based models such as BERT or GPT-2, so models with both image and language understanding abilities should be considered. We propose a new model that is effective for the ambiguous candidate identification task in DSTC11 SIMMC2.1 Tark. It consists of a simple pipeline model structure, which has two steps. The first step is to check whether there is ambiguity in the current user utterance, and the second step is to extract objects mentioned in the ambiguous utterance of the user. We suggest a new learning framework with a pre-trained image model and text model that is effective for the ambiguous candidate identification task. Experiments show that the proposed method can improve the model performance, and our model achieved 3rd place in sub-task 1 of the SIMMC2.1 track.

To fulfill complex user requirements in a situated conversational scenario, the agent needs to conduct step-by-step multi-modal logic reasoning, which includes locating objects, querying information and searching objects. However, existing methods omit this multi-step procedure and therefore constitutes the risk of shortcuts when making predictions. For example, they may directly copy the information from the dialogue history or simply use the textual description without perform visual reasoning. To address this issue and further boost the system performance, we apply the dual process theory to plug a reasoner into the original transformer based model for step-by-step reasoning. When system 2 completes multi-step reasoning, its output is regarded as final prediction. Our proposed method achieved the 1st rank on the summing scores across all four DSTC-11 SIMMC 2.1 sub-tasks.

The Situated Interactive MultiModal Conversations (SIMMC2.1) Challenge 2022 is hosted by the Eleventh Dialog System Technology Challenge (DSTC11). This is the third consecutive year multimodal dialog systems have been selected as an official track of the competition, promoted by the continued interest in the research community. The task of SIMMC is to create a shopping assistant agent that can communicate with customers in a virtual store. It requires processing store scenes and product catalogs along with the customer’s request. The task is decomposed into four steps and each becomes a subtask. In this work, we explore the common approaches to modeling multimodality and find the method with the most potential. We also identify a discrepancy in using pretrained language models for dialog tasks and devise a simple domain-adaptation method. Our model came in third place for object coreferencing, dialog state tracking, and response generation tasks.

Intent recognition is critical for task-oriented dialogue systems. However, for emerging domains and new services, it is difficult to accurately identify the key intent of a conversation due to time-consuming data annotation and comparatively poor model transferability. Therefore, the automatic induction of dialogue intention is very important for intelligent dialogue systems. This paper presents our solution to Track 2 of Intent Induction from Conversations for Task-Oriented Dialogue at the Eleventh Dialogue System Technology Challenge (DSTC11). The essence of intention clustering lies in distinguishing the representation of different dialogue utterances. The key to automatic intention induction is that, for any given set of new data, the sentence representation obtained by the model can be well distinguished from different labels. Therefore, we propose a multi-stage coarse-to-fine contrastive learning model training scheme including unsupervised contrastive learning pre-training, supervised contrastive learning pre-training, and fine-tuning with joint contrastive learning and clustering to obtain a better dialogue utterance representation model for the clustering task. In the released DSTC11 Track 2 evaluation results, our proposed system ranked first on both of the two subtasks of this Track.

We present our work on Track 2 in the Dialog System Technology Challenges 11 (DSTC11). DSTC11-Track2 aims to provide a benchmark for zero-shot, cross-domain, intent-set induction. In the absence of in-domain training dataset, robust utterance representation that can be used across domains is necessary to induce users’ intentions. To achieve this, we leveraged a multi-domain dialogue dataset to fine-tune the language model and proposed extracting Verb-Object pairs to remove the artifacts of unnecessary information. Furthermore, we devised the method that generates each cluster’s name for the explainability of clustered results. Our approach achieved 3rd place in the precision score and showed superior accuracy and normalized mutual information (NMI) score than the baseline model on various domain datasets.

Natural Language Understanding (NLU) is one of the most critical components of task-oriented dialogue, and it is often considered as an intent classification task. To achieve outstanding intent identification performance, system designers often need to hire a large number of domain experts to label the data, which is inefficient and costly. To address this problem, researchers’ attention has gradually shifted to automatic intent clustering methods, which employ low-resource unsupervised approaches to solve classification problems. The classical framework for clustering is deep clustering, which uses deep neural networks (DNNs) to jointly optimize non-clustering loss and clustering loss. However, for new conversational domains or services, utterances required to assign intents are scarce and the performance of DNNs is often dependent on large amounts of data. In addition, although re-clustering with k-means algorithm after training the network usually leads to better results, k-means methods often suffer from poor stability. To address these problems, we propose an effective two-stage progressive approach to refine the clustering. Firstly, we pre-train the network with contrastive loss using all conversations data and then optimize the clustering loss and contrastive loss simultaneously. Secondly, we propose adaptive progressive k-means to alleviate the randomness of vanilla k-means, achieving better performance and smaller deviation. Our method ranks second in DSTC11 Track2 Task 1, a benchmark for intent clustering of task-oriented dialogue, demonstrating the superiority and effectiveness of our method.

The focus of this work is to investigate unsupervised approaches to overcome quintessential challenges in designing task-oriented dialog schema: assigning intent labels to each dialog turn (intent clustering) and generating a set of intents based on the intent clustering methods (intent induction). We postulate there are two salient factors for automatic induction of intents: (1) clustering algorithm for intent labeling and (2) user utterance embedding space. We compare existing off-the-shelf clustering models and embeddings based on DSTC11 evaluation. Our extensive experiments demonstrate that the combined selection of utterance embedding and clustering method in the intent induction task should be carefully considered. We also present that pretrained MiniLM with Agglomerative clustering shows significant improvement in NMI, ARI, F1, accuracy and example coverage in intent induction tasks. The source codes are available at https://github.com/Jeiyoon/dstc11-track2.

In Task Oriented Dialogue (TOD) system, detecting and inducing new intents are two main challenges to apply the system in the real world. In this paper, we suggest the semantic multiview model to resolve these two challenges: (1) SBERT for General Embedding (GE), (2) Multi Domain Batch (MDB) for dialogue domain knowledge, and (3) Proxy Gradient Transfer (PGT) for cluster-specialized semantic. MDB feeds diverse dialogue datasets to the model at once to tackle the multi-domain problem by learning the multiple domain knowledge. We introduce a novel method PGT, which employs the Siamese network to fine-tune the model with a clustering method directly. Our model can learn how to cluster dialogue utterances by using PGT. Experimental results demonstrate that our multi-view model with MDB and PGT significantly improves the Open Intent Induction performance compared to baseline systems.

Although there have been remarkable advances in dialogue systems through the dialogue systems technology competition (DSTC), it remains one of the key challenges to building a robust task-oriented dialogue system with a speech interface. Most of the progress has been made for text-based dialogue systems since there are abundant datasets with written cor- pora while those with spoken dialogues are very scarce. However, as can be seen from voice assistant systems such as Siri and Alexa, it is of practical importance to transfer the success to spoken dialogues. In this paper, we describe our engineering effort in building a highly successful model that participated in the speech-aware dialogue systems technology challenge track in DSTC11. Our model consists of three major modules: (1) automatic speech recognition error correction to bridge the gap between the spoken and the text utterances, (2) text-based dialogue system (D3ST) for estimating the slots and values using slot descriptions, and (3) post-processing for recovering the error of the estimated slot value. Our experiments show that it is important to use an explicit automatic speech recognition error correction module, post-processing, and data augmentation to adapt a text-based dialogue state tracker for spoken dialogue corpora.

In a real-world environment, Dialogue State Tracking (DST) should use speech recognition results to perform tasks. However, most existing DST research has been conducted in text-based environments. This study aims to build a model that efficiently performs Automatic Speech Recognition-based DST. To operate robustly against speech noise, we used CopyT5, which adopted a copy mechanism, and trained the model using augmented data including speech noise. Furthermore, CopyT5 performed post-training using the masked language modeling method with the MultiWOZ dataset in T5 in order to learn the dialogue context better. The copy mechanism also mitigated name entity errors that may occur during DST generation. Experiments confirmed that data augmentation, post-training, and the copy mechanism effectively improve DST performance.

Though Dialogue State Tracking (DST) is a core component of spoken dialogue systems, recent work on this task mostly deals with chat corpora, disregarding the discrepancies between spoken and written language. In this paper, we propose OLISIA, a cascade system which integrates an Automatic Speech Recognition (ASR) model and a DST model. We introduce several adaptations in the ASR and DST modules to improve integration and robustness to spoken conversations. With these adaptations, our system ranked first in DSTC11 Track 3, a benchmark to evaluate spoken DST. We conduct an in-depth analysis of the results and find that normalizing the ASR outputs and adapting the DST inputs through data augmentation, along with increasing the pre-trained models size all play an important role in reducing the performance discrepancy between written and spoken conversations.

Prior research on dialogue state tracking (DST) is mostly based on written dialogue corpora. For spoken dialogues, the DST model trained on the written text should use the results (or hypothesis) of automatic speech recognition (ASR) as input. But ASR hypothesis often includes errors, which leads to significant performance drop for spoken dialogue state tracking. We address the issue by developing the following ASR error correction modules. First, we train a model to convert ASR hypothesis to ground truth user utterance, which can fix frequent patterns of errors. The model takes ASR hypotheses of two ASR models as input and fine-tuned in two stages. The corrected hypothesis is fed into a large scale pre-trained encoder-decoder model (T5) for DST training and inference. Second, if an output slot value from the encoder-decoder model is a name, we compare it with names in a dictionary crawled from Web sites and, if feasible, replace with the crawled name of the shortest edit distance. Third, we fix errors of temporal expressions in ASR hypothesis by using hand-crafted rules. Experiment results on the DSTC 11 speech-aware dataset, which is built on the popular MultiWOZ task (version 2.1), show that our proposed method can effectively mitigate the performance drop when moving from written text to spoken conversations.

This paper describes the systems submitted by team6 for ChatEval, the DSTC 11 Track 4 competition. We present three different approaches to predicting turn-level qualities of chatbot responses based on large language models (LLMs). We report improvement over the baseline using dynamic few-shot examples from a vector store for the prompts for ChatGPT. We also analyze the performance of the other two approaches and report needed improvements for future work. We developed the three systems over just two weeks, showing the potential of LLMs for this task. An ablation study conducted after the challenge deadline shows that the new Llama 2 models are closing the performance gap between ChatGPT and open-source LLMs. However, we find that the Llama 2 models do not benefit from few-shot examples in the same way as ChatGPT.

Open-domain automatic dialogue evaluation plays an important role in dialogue systems. While recent efforts are being put into making learning-based evaluation metrics correlate better with human evaluation, robust metrics for parallel corpora and multiple domains remain unexplored. Parallel corpora refer to corpora that express the same idea in different ways (e.g., translation, paraphrasing and back-translation). In this paper, we propose Parallel Corpora Alignment Framework (PCAF), which improves the consistency and robustness of model evaluation on parallel corpora. Firstly, parallel corpora are aligned in semantic space through parallel-corpora-aligned contrastive learning. Then, parallel-corpora-aligned distillation on multi-dataset is applied to further improve model’s generalization ability across multiple data domains. Our approach ranks second on the final test data of DSTC11 track4 subtask1 (“Multilingual Automatic Evaluation Metrics”, turn-level) and third on the subtask2 (“Robust Automatic Evaluation Metrics”, turn-level), which proves the strong generalization ability and robustness of our proposed approach.

Despite significant research effort in the development of automatic dialogue evaluation metrics, little thought is given to evaluating dialogues other than in English. At the same time, ensuring metrics are invariant to semantically similar responses is also an overlooked topic. In order to achieve the desired properties of robustness and multilinguality for dialogue evaluation metrics, we propose a novel framework that takes advantage of the strengths of current evaluation models with the newly-established paradigm of prompting Large Language Models (LLMs). Empirical results show our framework achieves state of the art results in terms of mean Spearman correlation scores across several benchmarks and ranks first place on both the Robust and Multilingual tasks of the DSTC11 Track 4 “Automatic Evaluation Metrics for Open-Domain Dialogue Systems”, proving the evaluation capabilities of prompted LLMs.

Task-oriented dialogue systems that employ external knowledge to generate informative responses have become an important field of research. This paper outlines our contribution to Track 5 of the Eleventh Dialog System Technology Challenge (DSTC11), which focuses on constructing high-performing, subjective knowledge-enriched task-oriented dialogue systems. Specifically, we investigate the complementarity of various language models to tackle the diverse knowledge selection task that involves multiple external sources. Based on this investigation, we propose pre- and post-generation model ensemble approaches to mitigate potential biases inherent in using a single model for the knowledge selection task. Finally, we utilize the consensus decoding approach to combine fine-tuned ensemble models and improve the performance of the generation system. Our system ranked 1st in human evaluation, even outperforming human annotation.

As Task-Oriented Dialog (TOD) systems have advanced, structured DB systems, which aim to collect relevant knowledge for answering user’s questions, have also progressed. Despite these advancements, these methods face challenges when dealing with subjective questions from users. To overcome this, DSTC11 released a subjective-knowledge-based TOD (SK-TOD) dataset and benchmark. This paper introduces a framework that effectively solves SK-TOD tasks by leveraging a Large Language Model (LLM). We demonstrate the proficient use of LLM for each sub-task, including an adapters-based method and knowledge-grounded data augmentation. Our proposed methods, which utilize LLM as an efficient tool, outperform baseline performance and approaches that directly use LLM as a one-step sub-task solver, showing superior task-specific optimization.

This paper presents our approach to building a generalized model for Track 5 in DSTC11: “Task-oriented Conversational Modeling with Subjective Knowledge” which addresses the challenge of generating responses to users’ utterances based on a variety of factual and subjective knowledge. To tackle this challenge, we first augmented the training data by leveraging contextual word embedding and back translation, thereby increasing the quantity of available data. Then, we utilized a large-size language model to enhance the acceptability of the augmented data and fine-tuned the model using augmented data. Specifically, we applied the DeBERTa-v3-large model for knowledge detection and selection, and the BART-large model for response generation. Our best model achieved the seventh rank in the objective evaluation and the second rank in the final official human evaluation. These outcomes serve as solid evidence that data augmentation and using a large-size model were highly effective for developing a conversational model system that incorporates objective and subjective knowledge.

This paper describes our submission to the fifth track of the 11th Dialog System Technology Challenge (DSTC-11), which focuses on “Task-oriented Conversational Modeling with Subjective Knowledge”. We focus on response generation and leverage a ranking strategy to ensemble individual models of BART, Long-T5, and a fine-tuned large language model based on LLaMA. The strategy is supplemented by other techniques like low rank adaptation to maintain efficient utilization of these large models while still achieving optimal performance. The experiments show that the ensemble method outperforms individual models and the baseline method. Our model was ranked 1st place in ROUGE_1, 2nd place in ROUGE_L score and 4th place in human evaluation among a total of 14 participating teams.

This paper presents our approach to the DSTC11 Track 5 selection task, which focuses on retrieving appropriate natural language knowledge sources for task-oriented dialogue. We propose typologically diverse back-translation method with typo noise, which could generate various structured user inquries. Through our noised back translation, we augmented inquiries by combining three different typologies of language sources with five different typo noise injections. Our experiments demonstrate that typological variety and typo noise aids the model in generalizing to diverse user inquiries in dialogue. In the competition, where 14 teams participated, our approach achieved the 5th rank for exact matching metric.

This paper discusses our approaches for task-oriented conversational modelling using subjective knowledge, with a particular emphasis on response generation. Our methodology was shaped by an extensive data analysis that evaluated key factors such as response length, sentiment, and dialogue acts present in the provided dataset. We used few-shot learning to augment the data with newly generated subjective knowledge items and present three approaches for DSTC11: (1) task-specific model exploration, (2) incorporation of the most frequent question into all generated responses, and (3) a waterfall prompting technique using a combination of both GPT-3 and ChatGPT.

The goal of DSTC11 track 5 is to build task-oriented dialogue systems that can effectively utilize external knowledge sources such as FAQs and reviews. This year’s challenge differs from previous ones as it includes subjective knowledge snippets and requires multiple snippets for a single turn. We propose a pipeline system for the challenge focusing on entity tracking, knowledge selection and response generation. Specifically, we devise a novel heuristic to ensemble the outputs from the rule-based method and neural model for entity tracking and knowledge selection. We also leverage metadata information in the knowledge source to handle fine-grained user queries. Our approach achieved the first place in objective evaluation and the third place in human evaluation of DSTC11 track 5.

We participate in the 11th Dialog System Technology Challenges (DSTC) track-5 called Task-oriented Conversational Modeling with Subjective Knowledge. Introducing subjective knowledge into task-oriented dialogue (TOD) can help the DS to understand variables of subjective user needs and to suit more dialogue scenarios. Track-5 includes several sub-tasks: 1) knowledge-seeking turn detection; 2) knowledge entity tracking; 3) knowledge entry selection; and 4) use of the selected knowledge entries for response generation. Besides the challenges of each sub-tasks own, there are two challenges across different sub-tasks. The first is that there are multiple valid knowledge entries for each knowledge-seeking turn, the accuracy of the knowledge entry selection is important for the quality of response generation. The second challenge is how to address the unseen dialogue/entities/entries in the validation and the test set. In this paper, we propose a difference-aware ensemble method to address these sub-tasks and the two challenges mentioned above. Our method helps to obtain more robust results and performs well on unseen instances. Among all the submissions for the test set, our method ranks 1st on the knowledge-seeking turn detection task and achieves 3rd on the overall automatic evaluation score. Our code and data will be released on GitHub.

Most research on task oriented dialog modeling is based on written text input. However, users interact with practical dialog systems often using speech as input. Typically, systems convert speech into text using an Automatic Speech Recognition (ASR) system, introducing errors. Furthermore, these systems do not address the differences in written and spoken language. The research on this topic is stymied by the lack of a public corpus. Motivated by these considerations, our goal in hosting the speech-aware dialog state tracking challenge was to create a public corpus or task which can be used to investigate the performance gap between the written and spoken forms of input, develop models that could alleviate this gap, and establish whether Text-to-Speech-based (TTS) systems is a reasonable surrogate to the more-labor intensive human data collection. We created three spoken versions of the popular written-domain MultiWoz task – (a) TTS-Verbatim: written user inputs were converted into speech waveforms using a TTS system, (b) Human-Verbatim: humans spoke the user inputs verbatim, and (c) Human-paraphrased: humans paraphrased the user inputs. Additionally, we provided different forms of ASR output to encourage wider participation from teams that may not have access to state-of-the-art ASR systems. These included ASR transcripts, word time stamps, and latent representations of the audio (audio encoder outputs). In this paper, we describe the corpus, report results from participating teams, provide preliminary analyses of their results, and summarize the current state-of-the-art in this domain.

With ever increasing interest in task-oriented dialog systems, the recent work on Situated and Interactive Multimodal Conversations (SIMMC 2.0) aims to develop personal assistants that interact with users, grounded in an immersive and co-observed setting of photo-realistic scenes. The dataset contains 11k task-oriented dialogs set in an interactive shopping scenario, spanning more than 117k utterances. In order to push research towards this next generation virtual assistants, the SIMMC 2.1 challenge was conducted at the Eleventh Dialog System Technology Challenge (DSTC) which had entries from across the world competing to achieve the state-of-the-art performance in the SIMMC 2.1 task. In this report, we present and compare 13 SIMMC 2.1 model entries from 5 trams across the world to understand the current progress made across the last three years (starting with SIMMC 1.0 and 2.0 challenges) for multimodal task-oriented dialog systems. We hope that our analysis throws light on components that showed promise in addition to identifying the gaps for future research towards this grand goal of an immersive multimodal conversational agent.

With increasing demand for and adoption of virtual assistants, recent work has investigated ways to accelerate bot schema design through the automatic induction of intents or the induction of slots and dialogue states. However, a lack of dedicated benchmarks and standardized evaluation has made progress difficult to track and comparisons between systems difficult to make. This challenge track, held as part of the Eleventh Dialog Systems Technology Challenge, introduces a benchmark that aims to evaluate methods for the automatic induction of customer intents in a realistic setting of customer service interactions between human agents and customers. We propose two subtasks for progressively tackling the automatic induction of intents and corresponding evaluation methodologies. We then present three datasets suitable for evaluating the tasks and propose simple baselines. Finally, we summarize the submissions and results of the challenge track, for which we received submissions from 34 teams.

The advent and fast development of neural networks have revolutionized the research on dialogue systems and subsequently have triggered various challenges regarding their automatic evaluation. Automatic evaluation of open-domain dialogue systems as an open challenge has been the center of the attention of many researchers. Despite the consistent efforts to improve automatic metrics’ correlations with human evaluation, there have been very few attempts to assess their robustness over multiple domains and dimensions. Also, their focus is mainly on the English language. All of these challenges prompt the development of automatic evaluation metrics that are reliable in various domains, dimensions, and languages. This track in the 11th Dialogue System Technology Challenge (DSTC11) is part of the ongoing effort to promote robust and multilingual automatic evaluation metrics. This article describes the datasets and baselines provided to participants and discusses the submission and result details of the two proposed subtasks.

Conventional Task-oriented Dialogue (TOD) Systems rely on domain-specific APIs/DBs or external factual knowledge to create responses. In DSTC11 track 5, we aims to provide a new challenging task to accommodate subjective user requests (e.g.,”Is the WIFI reliable?” or “Does the restaurant have a good atmosphere?” into TOD. We release a benchmark dataset, which contains subjective knowledge-seeking dialogue contexts and manually annotated responses that are grounded in subjective knowledge sources. The challenge track received a total of 48 entries from 14 participating teams.

Explanations accompanying a recommendation can assist users in understanding the decision made by recommendation systems, which in turn increases a user’s confidence and trust in the system. Recently, research has focused on generating natural language explanations in a human-readable format. Thus far, the proposed approaches leverage item reviews written by users, which are often subjective, sparse in language, and unable to account for new items that have not been purchased or reviewed before. Instead, we aim to generate fact-grounded recommendation explanations that are objectively described with item features while implicitly considering a user’s preferences, based on the user’s purchase history. To achieve this, we propose a knowledge graph (KG) approach to natural language explainable recommendation. Our approach draws on user-item features through a novel collaborative filtering-based KG representation to produce fact-grounded, personalized explanations, while jointly learning user-item representations for recommendation scoring. Experimental results show that our approach consistently outperforms previous state-of-the-art models on natural language explainable recommendation metrics.

How do sequence models represent their decision-making process? Prior work suggests that Othello-playing neural network learned nonlinear models of the board state (Li et al., 2023a). In this work, we provide evidence of a closely related linear representation of the board. In particular, we show that probing for “my colour” vs. “opponent’s colour” may be a simple yet powerful way to interpret the model’s internal state. This precise understanding of the internal representations allows us to control the model’s behaviour with simple vector arithmetic. Linear representations enable significant interpretability progress, which we demonstrate with further exploration of how the world model is computed.

Large language models (LLMs) have displayed an impressive ability to harness natural language to perform complex tasks. We explore whether we can leverage this ability to find and explain patterns in data. Specifically, given a pre-trained LLM and data examples, we apply interpretable autoprompting (iPrompt) to generate a natural language string explaining the data. iPrompt iteratively generates explanations with an LLM and reranks them based on their performance when used as a prompt. Experiments on a wide range of datasets, from synthetic mathematics to natural language understanding, show that iPrompt can yield meaningful insights by accurately finding dataset explanations that are human-interpretable. Moreover, iPrompt is reasonably efficient, as it does not require access to model gradients and works with relatively small models (e.g. ~6 billion parameters rather than >=100 billion). Finally, experiments with scientific datasets show the potential for iPrompt to aid in scientific discovery.

With their increasing size, large language models (LLMs) are becoming increasingly good at language understanding tasks. But even with high performance on specific downstream task, LLMs fail at simple linguistic tests for negation or quantifier understanding. Previous work on quantifier understanding in LLMs show inverse scaling in understanding few-type quantifiers. In this paper, we question the claims of of previous work and show that it is a result of inappropriate testing methodology. We also present alternate methods to measure quantifier comprehension in LLMs and show that LLMs are able to better understand the difference between the meaning of few-type and most-type quantifiers as their size increases, although they are not particularly good at it. We also observe inverse scaling for most-type quantifier understanding, which is contrary to human psycho-linguistic experiments and previous work, where the model’s understanding of most-type quantifier gets worse as the model size increases. We do this evaluation on models ranging from 125M-175B parameters, which suggests that LLMs do not do as well as expected with quantifiers. We also discuss the possible reasons for this and the relevance of quantifier understanding in evaluating language understanding in LLMs.

Differential Privacy (DP) has been tailored to address the unique challenges of text-to-text privatization. However, text-to-text privatization is known for degrading the performance of language models when trained on perturbed text. Employing a series of interpretation techniques on the internal representations extracted from BERT trained on perturbed pre-text, we intend to disentangle at the linguistic level the distortion induced by differential privacy. Experimental results from a representational similarity analysis indicate that the overall similarity of internal representations is substantially reduced. Using probing tasks to unpack this dissimilarity, we find evidence that text-to-text privatization affects the linguistic competence across several formalisms, encoding localized properties of words while falling short at encoding the contextual relationships between spans of words.

The ubiquity of complex machine learning has raised the importance of model-agnostic explanation algorithms. These methods create artificial instances by slightly perturbing real instances, capturing shifts in model decisions. However, such methods rely on initial data and only provide explanations of the decision for these. To tackle these problems, we propose Therapy, the first global and model-agnostic explanation method adapted to text which requires no input dataset. Therapy generates texts following the distribution learned by a classifier through cooperative generation. Because it does not rely on initial samples, it allows to generate explanations even when data is absent (e.g., for confidentiality reasons). Moreover, conversely to existing methods that combine multiple local explanations into a global one, Therapy offers a global overview of the model behavior on the input space. Our experiments show that although using no input data to generate samples, Therapy provides insightful information about features used by the classifier that is competitive with the ones from methods relying on input samples and outperforms them when input samples are not specific to the studied model.

Large language models (LLMs) that do not give consistent answers across contexts are problematic when used for tasks with expectations of consistency–e.g. question-answering, explanations, etc. Our work presents an evaluation benchmark for self-consistency in cases of under-specification where two or more answers can be correct. We conduct a series of behavioral experiments on the OpenAI model suite using an ambiguous integer sequence completion task. We find that average consistency ranges from 67% to 82%, far higher than would be predicted if a model’s consistency was random, and increases as model capability improves. Furthermore, we show that models tend to maintain self-consistency across a series of robustness checks, including prompting speaker changes and sequence length changes. These results suggest that self-consistency arises as an emergent capability without specifically training for it. Despite this, we find that models are uncalibrated when judging their own consistency, with models displaying both over- and under-confidence. We also propose a nonparametric test for determining from token output distribution whether a model assigns non-trivial probability to alternative answers. Using this test, we find that despite increases in self-consistency, models usually place significant weight on alternative, inconsistent answers. This distribution of probability mass provides evidence that even highly self-consistent models internally compute multiple possible responses.

The Backpack is a Transformer alternative shown to improve interpretability in English language modeling by decomposing predictions into a weighted sum of token sense components. However, Backpacks’ reliance on token-defined meaning raises questions as to their potential for languages other than English, a language for which subword tokenization provides a reasonable approximation for lexical items. In this work, we train, evaluate, interpret, and control Backpack language models in character-tokenized Chinese, in which words are often composed of many characters. We find that our (134M parameter) Chinese Backpack language model performs comparably to a (104M parameter) Transformer, and learns rich character-level meanings that log-additively compose to form word meanings. In SimLex-style lexical semantic evaluations, simple averages of Backpack character senses outperform input embeddings from a Transformer. We find that complex multi-character meanings are often formed by using the same per-character sense weights consistently across context. Exploring interpretability-through control, we show that we can localize a source of gender bias in our Backpacks to specific character senses and intervene to reduce the bias.

Recent research suggests that the feed-forward module within Transformers can be viewed as a collection of key-value memories, where the keys learn to capture specific patterns from the input based on the training examples. The values then combine the output from the ‘memories’ of the keys to generate predictions about the next token. This leads to an incremental process of prediction that gradually converges towards the final token choice near the output layers. This interesting perspective raises questions about how multilingual models might leverage this mechanism. Specifically, for autoregressive models trained on two or more languages, do all neurons (across layers) respond equally to all languages? No! Our hypothesis centers around the notion that during pre-training, certain model parameters learn strong language-specific features, while others learn more language-agnostic (shared across languages) features. To validate this, we conduct experiments utilizing parallel corpora of two languages that the model was initially pre-trained on. Our findings reveal that the layers closest to the network’s input or output tend to exhibit more language-specific behaviour compared to the layers in the middle.

A recent body of work has demonstrated that Transformer embeddings can be linearly decomposed into well-defined sums of factors, that can in turn be related to specific network inputs or components. There is however still a dearth of work studying whether these mathematical reformulations are empirically meaningful. In the present work, we study representations from machine-translation decoders using two of such embedding decomposition methods. Our results indicate that, while decomposition-derived indicators effectively correlate with model performance, variation across different runs suggests a more nuanced take on this question. The high variability of our measurements indicate that geometry reflects model-specific characteristics more than it does sentence-specific computations, and that similar training conditions do not guarantee similar vector spaces.

The question of what kinds of linguistic information are encoded in different layers of Transformer-based language models is of considerable interest for the NLP community. Existing work, however, has overwhelmingly focused on word-level representations and encoder-only language models with the masked-token training objective. In this paper, we present experiments with semantic structural probing, a method for studying sentence-level representations via finding a subspace of the embedding space that provides suitable task-specific pairwise distances between data-points. We apply our method to language models from different families (encoder-only, decoder-only, encoder-decoder) and of different sizes in the context of two tasks, semantic textual similarity and natural-language inference. We find that model families differ substantially in their performance and layer dynamics, but that the results are largely model-size invariant.

Recent work suggests that large language models (LLMs) achieve higher accuracy on multi-step reasoning tasks when prompted to generate intermediate reasoning steps, or a chain of thought (CoT), before their final answer. However, it is unclear how exactly CoTs improve LLMs’ accuracy, and in particular, if LLMs use their CoTs to reason to their final answers. This paper tries to answer this question with respect to arithmetic word problems, by (i) evaluating the correctness of LLMs’ CoTs, and (ii) using causal abstraction to assess if the intermediate tokens produced as part of a CoT causally impact LLMs’ final answers, in line with the reasoning described by the CoT. We find that for CoT-prompted LLMs, correct answers to arithmetic problems are highly correlated with correct CoTs, and that when LLMs produce correct CoTs, they realize to a fairly large extent the causal models suggested by their CoTs. Higher degrees of realization also seem associated with better overall accuracy on the arithmetic problems. These findings suggest that some CoT-prompted LLMs may do better on multi-step arithmetic reasoning at least partly because they use their CoTs to reason to their final answers. However, for some LLMs, other internal processes may also be involved.

Interpretability methods in NLP aim to provide insights into the semantics underlying specific system architectures. Focusing on word embeddings, we present a supervised-learning method that, for a given domain (e.g., sports, professions), identifies a subset of model features that strongly improve prediction of human similarity judgments. We show this method keeps only 20-40% of the original embeddings, for 8 independent semantic domains, and that it retains different feature sets across domains. We then present two approaches for interpreting the semantics of the retained features. The first obtains the scores of the domain words (co-hyponyms) on the first principal component of the retained embeddings, and extracts terms whose co-occurrence with the co-hyponyms tracks these scores’ profile. This analysis reveals that humans differentiate e.g. sports based on how gender-inclusive and international they are. The second approach uses the retained sets as variables in a probing task that predicts values along 65 semantically annotated dimensions for a dataset of 535 words. The features retained for professions are best at predicting cognitive, emotional and social dimensions, whereas features retained for fruits or vegetables best predict the gustation (taste) dimension. We discuss implications for alignment between AI systems and human knowledge.

The increasing interest in probing the linguistic capabilities of large language models (LLMs) has long reached the area of semantics and pragmatics, including the phenomenon of presuppositions. In this study, we investigate a phenomenon that, however, has not yet been investigated, i.e., the phenomenon of anti-presupposition and the principle that accounts for it, the Maximize Presupposition! principle (MP!). Through an experimental investigation using psycholinguistic data and four open-source BERT model variants, we explore how language models handle different anti-presuppositions and whether they apply the MP! principle in their predictions. Further, we examine whether fine-tuning with Natural Language Inference data impacts adherence to the MP! principle. Our findings reveal that LLMs tend to replicate context-based n-grams rather than follow the MP! principle, with fine-tuning not enhancing their adherence. Notably, our results further indicate a striking difficulty of LLMs to correctly predict determiners, in relatively simple linguistic contexts.

Examples are a powerful tool that help us understand complex concepts and connections. In computational linguistics research, looking at example system output and example corpus entries can offer a wealth of insights that are not otherwise accessible. This paper describes the open-source software VULCAN, a visualization tool for strings, graphs, trees, alignments, attention and more. VULCAN’s unique ability to visualize both linguistic structures and properties of neural models make it particularly relevant for neuro-symbolic models. Neuro-symbolic models, combining neural networks with often linguistically grounded structures, offer a promise of increased interpretability in an age of purely neural black-box end-to-end models. VULCAN aims to facilitate this interpretability in practice. VULCAN is designed to be both easy to use and powerful in its capabilities.

Several methodologies have recently been proposed to evaluate the ability of Pretrained Language Models (PLMs) to interpret negation. In this article, we build on Gubelmann and Handschuh (2022), which studies the modification of PLMs’ predictions as a function of the polarity of inputs, in English. Crucially, this test uses “self-contained” inputs ending with a masked position: depending on the polarity of a verb in the input, a particular token is either semantically ruled out or allowed at the masked position. By replicating Gubelmann and Handschuh (2022) experiments, we have uncovered flaws that weaken the conclusions that can be drawn from this test. We thus propose an improved version, the Self-Contained Neg Test, which is more controlled, more systematic, and entirely based on examples forming minimal pairs varying only in the presence or absence of verbal negation in English. When applying our test to the roberta and bert base and large models, we show that only roberta-large shows trends that match the expectations, while bert-base is mostly insensitive to negation. For all the tested models though, in a significant number of test instances the top-1 prediction remains the token that is semantically forbidden by the context, which shows how much room for improvement remains for a proper treatment of the negation phenomenon.

As neural language models (NLMs) based on Transformers are becoming increasingly dominant in natural language processing, several studies have proposed analyzing the semantic and pragmatic abilities of such models. In our study, we aimed at investigating the effect of discourse connectives on NLMs with regard to Transformer Surprisal scores by focusing on the English stimuli of an experimental dataset, in which the expectations about an event in a discourse fragment could be reversed by a concessive or a contrastive connective. By comparing the Surprisal scores of several NLMs, we found that bigger NLMs show patterns similar to humans’ behavioral data when a concessive connective is used, while connective-related effects tend to disappear with a contrastive one. We have additionally validated our findings with GPT-Neo using an extended dataset, and results mostly show a consistent pattern.

Probing contextualized representations typically involves comparing task-specific model predictions against ground truth linguistic labels. Although this methodology shows what information can be recovered by a classifier, it does not reveal how a classifier uses the representation to make its decision. To address the latter problem, we ask: Do task-classifiers rely on representation- and task-independent geometric patterns in the embedding space? We explore this question by developing MetaProbe, an approach that uses geometric properties of representations to predict the behavior of task-specific classifiers (i.e., their predictions as opposed to the ground truth). Our experiments reveal the existence of universal geometric patterns across representations that can predict classifier predictions. Consequently, this allows us to posit a geometric explanation for the impressive performance of contextualized representations.

Contrast set consistency is a robustness measurement that evaluates the rate at which a model correctly responds to all instances in a bundle of minimally different examples relying on the same knowledge. To draw additional insights, we propose to complement consistency with relative consistency—the probability that an equally accurate model would surpass the consistency of the proposed model, given a distribution over possible consistencies. Models with 100% relative consistency have reached a consistency peak for their accuracy. We reflect on prior work that reports consistency in contrast sets and observe that relative consistency can alter the assessment of a model’s consistency compared to another. We anticipate that our proposed measurement and insights will influence future studies aiming to promote consistent behavior in models.

Pretrained language models (PLMs) form the basis of most state-of-the-art NLP technologies. Nevertheless, they are essentially black boxes: Humans do not have a clear understanding of what knowledge is encoded in different parts of the models, especially in individual neurons. A contrast is in computer vision, where feature visualization provides a decompositional interpretability technique for neurons of vision models. Activation maximization is used to synthesize inherently interpretable visual representations of the information encoded in individual neurons. Our work is inspired by this but presents a cautionary tale on the interpretability of single neurons, based on the first large-scale attempt to adapt activation maximization to NLP, and, more specifically, large PLMs. We propose feature textualization, a technique to produce dense representations of neurons in the PLM word embedding space. We apply feature textualization to the BERT model to investigate whether the knowledge encoded in individual neurons can be interpreted and symbolized. We find that the produced representations can provide insights about the knowledge encoded in individual neurons, but that individual neurons do not represent clear-cut symbolic units of language such as words. Additionally, we use feature textualization to investigate how many neurons are needed to encode words in BERT.

Despite the fact that Transformers perform well in NLP tasks, recent studies suggest that self-attention is theoretically limited in learning even some regular and context-free languages. These findings motivated us to think about their implications in modeling natural language, which is hypothesized to be mildly context-sensitive. We test the Transformer’s ability to learn mildly context-sensitive languages of varying complexities, and find that they generalize well to unseen in-distribution data, but their ability to extrapolate to longer strings is worse than that of LSTMs. Our analyses show that the learned self-attention patterns and representations modeled dependency relations and demonstrated counting behavior, which may have helped the models solve the languages.

Large language models (LLMs) like GPT and PALM have excelled in numerous natural language processing (NLP) tasks such as text generation, question answering, and translation. However, they are also found to have inherent social biases. To address this, recent studies have proposed debiasing techniques like iterative nullspace projection (INLP) and Counterfactual Data Augmentation (CDA). Additionally, there’s growing interest in understanding the intricacies of these models. Some researchers focus on individual neural units, while others examine specific layers. In our study, we benchmark newly released models, assess the impact of debiasing methods, and investigate how biases are linked to different transformer layers using a method called Logit Lens. Specifically, we evaluate three modern LLMs: OPT, LLaMA, and LLaMA2, and their debiased versions. Our experiments are based on two popular bias evaluation datasets, StereoSet and CrowS-Pairs, and we perform a layer-by-layer analysis using the Logit Lens.

Vector-space models of word meaning all assume that words occurring in similar contexts have similar meanings. Words that are similar in their topical associations but differ in their logical force tend to emerge as semantically close – creating well-known challenges for NLP applications that involve logical reasoning. Pretrained language models such as BERT, RoBERTa, GPT-2, and GPT-3 hold the promise of performing better on logical tasks than classic static word embeddings. However, reports are mixed about their success. Here, we advance this discussion through a systematic study of scalar adverbs, an under-explored class of words with strong logical force. Using three different tasks involving both naturalistic social media data and constructed examples, we investigate the extent to which BERT, RoBERTa, GPT-2 and GPT-3 exhibit knowledge of these common words. We ask: 1) Do the models distinguish amongst the three semantic categories of MODALITY, FREQUENCY and DEGREE? 2) Do they have implicit representations of full scales from maximally negative to maximally positive? 3) How do word frequency and contextual factors impact model performance? We find that despite capturing some aspects of logical meaning, the models still have obvious shortfalls.

Natural language is an appealing medium for explaining how large language models process and store information, but evaluating the faithfulness of such explanations is challenging. To help address this, we develop two modes of evaluation for natural language explanations that claim individual neurons represent a concept in a text input. In the *observational mode*, we evaluate claims that a neuron a activates on all and only input strings that refer to a concept picked out by the proposed explanation E. In the *intervention mode*, we construe E as a claim that neuron a is a causal mediator of the concept denoted by E. We apply our framework to the GPT-4-generated explanations of GPT-2 XL neurons of Bills et al. (2023) and show that even the most confident explanations have high error rates and little to no causal efficacy. We close the paper by critically assessing whether natural language is a good choice for explanations and whether neurons are the best level of analysis.

We examine the licensing of negative polarity items (NPIs) in large language models (LLMs) to enrich the picture of how models acquire NPIs as linguistic phenomena at the syntax-semantics interface. NPIs are a class of words which have a restricted distribution, appearing only in certain licensing contexts, prototypically negation. Unlike much of previous work which assumes NPIs and their licensing environments constitute unified classes, we consider NPI distribution in its full complexity: different NPIs are possible in different licensing environments. By studying this phenomenon across a broad range of models, we are able to explore which features of the model architecture, properties of the training data, and linguistic characteristics of the NPI phenomenon itself drive performance.

Answering multi-hop reasoning questions requires retrieving and synthesizing information from diverse sources. Large Language Models (LLMs) struggle to perform such reasoning consistently. Here we propose an approach to pinpoint and rectify multi-hop reasoning failures through targeted memory injections on LLM attention heads. First, we analyze the per-layer activations of GPT-2 models in response to single and multi-hop prompts. We then propose a mechanism that allows users to inject pertinent prompt-specific information, which we refer to as “memories,” at critical LLM locations during inference. By thus enabling the LLM to incorporate additional relevant information during inference, we enhance the quality of multi-hop prompt completions. We show empirically that a simple, efficient, and targeted memory injection into a key attention layer can often increase the probability of the desired next token in multi-hop tasks, by up to 424%.

Constituents are groups of words that behave as a syntactic unit. Many linguistic phenomena (e.g., question formation, diathesis alternations) require the manipulation and rearrangement of constituents in a sentence. In this paper, we investigate how different finetuning setups affect the ability of pretrained sequence-to-sequence language models such as BART and T5 to replicate constituency tests — transformations that involve manipulating constituents in a sentence. We design multiple evaluation settings by varying the combinations of constituency tests and sentence types that a model is exposed to during finetuning. We show that models can replicate a linguistic transformation on a specific type of sentence that they saw during finetuning, but performance degrades substantially in other settings, showing a lack of systematic generalization. These results suggest that models often learn to manipulate sentences at a surface level unrelated to the constituent-level syntactic structure, for example by copying the first word of a sentence. These results may partially explain the brittleness of pretrained language models in downstream tasks.

Psycholinguistic studies suggested that our mental perception of events depends not only on the lexical items used to describe them, but also on the syntactic structure of the event description. More specifically, it has been argued that light verb constructions affect the perception of duration in event construal, such that the same event in this type of constructions is perceived by humans as taking less time (to give a kiss takes a shorter time than to kiss). In our paper, we present two experiments with BERT using English stimuli from psycholinguistic studies to investigate the effects of the syntactic construction on event duration and event similarity. We show that i) the dimensions of BERT vectors encode a smaller value for duration for both punctive and durative events in count syntax, in line with human results; on the other hand, we also found that ii) BERT semantic similarity fails to capture the conceptual shift that durative events should undergo in count syntax.

Language models (LMs) exhibit and amplify many types of undesirable biases learned from the training data, including gender bias. However, we lack tools for effectively and efficiently changing this behavior without hurting general language modeling performance. In this paper, we study three methods for identifying causal relations between LM components and particular output: causal mediation analysis, automated circuit discovery and our novel, efficient method called DiffMask+ based on differential masking. We apply the methods to GPT-2 small and the problem of gender bias, and use the discovered sets of components to perform parameter-efficient fine-tuning for bias mitigation. Our results show significant overlap in the identified components (despite huge differences in the computational requirements of the methods) as well as success in mitigating gender bias, with less damage to general language modeling compared to full model fine-tuning. However, our work also underscores the difficulty of defining and measuring bias, and the sensitivity of causal discovery procedures to dataset choice. We hope our work can contribute to more attention for dataset development, and lead to more effective mitigation strategies for other types of bias.

Anthropomorphization is the tendency to attribute human-like traits to non-human entities. It is prevalent in many social contexts – children anthropomorphize toys, adults do so with brands, and it is a literary device. It is also a versatile tool in science, with behavioral psychology and evolutionary biology meticulously documenting its consequences. With widespread adoption of AI systems, and the push from stakeholders to make it human-like through alignment techniques, human voice, and pictorial avatars, the tendency for users to anthropomorphize it increases significantly. We take a dyadic approach to understanding this phenomenon with large language models (LLMs) by studying (1) the objective legal implications, as analyzed through the lens of the recent blueprint of AI bill of rights and the (2) subtle psychological aspects customization and anthropomorphization. We find that anthropomorphized LLMs customized for different user bases violate multiple provisions in the legislative blueprint. In addition, we point out that anthropomorphization of LLMs affects the influence they can have on their users, thus having the potential to fundamentally change the nature of human-AI interaction, with potential for manipulation and negative influence. With LLMs being hyper-personalized for vulnerable groups like children and patients among others, our work is a timely and important contribution. We propose a conservative strategy for the cautious use of anthropomorphization to improve trustworthiness of AI systems.

The process of identifying obligations in a legal text is not a straightforward task, because not only are the documents long, but the sentences therein are long as well. As a result of long elements in the text, law is more difficult to interpret (Coupette et al., 2021). Moreover, the identification of obligations relies not only on the clarity and precision of the language used but also on the unique perspectives, experiences, and knowledge of the reader. In particular, this paper addresses the problem of identifyingobligations using machine and deep learning approaches showing a full comparison between both methodologies and proposing a new approach called NOMOS based on the combination of Positional Embeddings (PE) and Temporal Convolutional Networks (TCNs). Quantitative and qualitative experiments, conducted on legal regulations 1, demonstrate the effectiveness of the proposed approach.

In the legal domain, we often perform classification tasks on very long documents, for example court judgements. These documents often contain thousands of words, so the length of these documents poses a challenge for this modelling task. In this research paper, we present a comprehensive evaluation of various strategies to perform long text classification using Transformers in conjunction with strategies to select document chunks using traditional NLP models. We conduct our experiments on 6 benchmark datasets comprising lengthy documents, 4 of which are publicly available. Each dataset has a median word count exceeding 1,000. Our evaluation encompasses state-of-the-art Transformer models, such as RoBERTa, Longformer, HAT, MEGA and LegalBERT and compares them with a traditional baseline TF-IDF + Neural Network (NN) model. We investigate the effectiveness of pre-training on large corpora, fine tuning strategies, and transfer learning techniques in the context of long text classification.

Language Models (LMs) have proven their ability to acquire diverse linguistic knowledge during the pretraining phase, potentially serving as a valuable source of incidental supervision for downstream tasks. However, there has been limited research conducted on the retrieval of domain-specific knowledge, and specifically legal knowledge. We propose to explore the task of Entity Typing, serving as a proxy for evaluating legal knowledge as an essential aspect of text comprehension, and a foundational task to numerous downstream legal NLP applications. Through systematic evaluation and analysis and two types of prompting (cloze sentences and QA-based templates) and to clarify the nature of these acquired cues, we compare diverse types and lengths of entities both general and domain-specific entities, semantics or syntax signals, and different LM pretraining corpus (generic and legal-oriented) and architectures (encoder BERT-based and decoder-only with Llama2). We show that (1) Llama2 performs well on certain entities and exhibits potential for substantial improvement with optimized prompt templates, (2) law-oriented LMs show inconsistent performance, possibly due to variations in their training corpus, (3) LMs demonstrate the ability to type entities even in the case of multi-token entities, (4) all models struggle with entities belonging to sub-domains of the law (5) Llama2 appears to frequently overlook syntactic cues, a shortcoming less present in BERT-based architectures.

NLP systems are increasingly used in the law domain, either by legal institutions or by the industry. As a result there is a pressing need to characterize their strengths and weaknesses and understand their inner workings. This article presents a case study on the task of judicial decision prediction, on a small dataset from French Courts of Appeal. Specifically, our dataset of around 1000 decisions is about the habitual place of residency of children from divorced parents. The task consists in predicting, from the facts and reasons of the documents, whether the court rules that children should live with their mother or their father. Instead of feeding the whole document to a classifier, we carefully construct the dataset to make sure that the input to the classifier does not contain any ‘spoilers’ (it is often the case in court rulings that information all along the document mentions the final decision). Our results are mostly negative: even classifiers based on French pretrained language models (Flaubert, JuriBERT) do not classify the decisions with a reasonable accuracy. However, they can extract the decision when it is part of the input. With regards to these results, we argue that there is a strong caveat when constructing legal NLP datasets automatically.

This work introduces a novel, extensive annotated corpus for multi-label legislative text classification in Italian, based on legal acts from the Gazzetta Ufficiale, the official source of legislative information of the Italian state. The annotated dataset, which we released to the community, comprises over 363,000 titles of legislative acts, spanning over 30 years from 1988 until 2022. Moreover, we evaluate four models for text classification on the dataset, demonstrating how using only the acts’ titles can achieve top-level classification performance, with a micro F1-score of 0.87. Also, our analysis shows how Italian domain-adapted legal models do not outperform general-purpose models on the task. Models’ performance can be checked by users via a demonstrator system provided in support of this work.

The application of Natural Language Processing (NLP) to specialized domains, such as the law, has recently received a surge of interest. As many legal services rely on processing and analyzing large collections of documents, automating such tasks with NLP tools such as language models emerges as a key challenge since legal documents may contain specialized vocabulary from other domains, such as medical terminology in personal injury text. However, most language models are general-purpose models, which either have limited reasoning capabilities on highly specialized legal terminology and syntax, such as BERT or ROBERTA, or are expensive to run and tune, such as GPT-3.5 and Claude. Thus, in this paper, we propose a specialized language model for personal injury text, LEGALRELECTRA, which is trained on mixed-domain legal and medical corpora. We show that as a small language model, our model improves over general-domain and single-domain medical and legal language models when processing mixed-domain (personal injury) text. Our training architecture implements the ELECTRA framework but utilizes REFORMER instead of BERT for its generator and discriminator. We show that this improves the model’s performance on processing long passages and results in better long-range text comprehension.

Finding the answers to legal questions about specific clauses in contracts is an important analysis in many legal workflows (e.g., understanding market trends, due diligence, risk mitigation) but more important is being able to do this at scale. In this paper, we present an examination of using large language models to produce (partially) structured answers to legal questions; primarily in the form of multiple choice and multiple select. We first show that traditional semantic matching is unable to perform this task at acceptable accuracy and then show how question specific prompts can achieve reasonable accuracy across a range of generative models. Finally, we show that much of this effectiveness can be maintained when generalized prompt templates are used rather than question specific ones.

The field of Legal Judgment Prediction (LJP) has witnessed significant growth in the past decade, with over 100 papers published in the past three years alone. Our comprehensive survey of over 150 papers reveals a stark reality: only ~7% of published papers are doing what they set out to do - predict court decisions. We delve into the reasons behind the flawed and unreliable nature of the remaining experiments, emphasising their limited utility in the legal domain. We examine the distinctions between predicting court decisions and the practices of legal professionals in their daily work. We explore how a lack of attention to the identity and needs of end-users has fostered the misconception that LJP is a near-solved challenge suitable for practical application, and contributed to the surge in academic research in the field. To address these issues, we examine three different dimensions of ‘doing LJP right’: using data appropriate for the task; tackling explainability; and adopting an application-centric approach to model reporting and evaluation. We formulate a practical checklist of recommendations, delineating the characteristics that are required if a judgment prediction system is to be a valuable addition to the legal field.

This paper formulates a new task of extracting privacy parameters from a privacy policy, through the lens of Contextual Integrity (CI), an established social theory framework for reasoning about privacy norms. Through extensive experiments, we further show that incorporating CI-based domain-specific knowledge into a BERT-based SRL model results in the highest precision and recall, achieving an F1 score of 84%. With our work, we would like to motivate new research in building NLP applications for the privacy domain.

Commercial contracts are known to be a valuable source for deriving project-specific requirements. However, contract negotiations mainly occur among the legal counsel of the parties involved. The participation of non-legal stakeholders, including requirement analysts, engineers, and solution architects, whose primary responsibility lies in ensuring the seamless implementation of contractual terms, is often indirect and inadequate. Consequently, a significant number of sentences in contractual clauses, though legally accurate, can appear unfair from an implementation perspective to non-legal stakeholders. This perception poses a problem since requirements indicated in the clauses are obligatory and can involve punitive measures and penalties if not implemented as committed in the contract. Therefore, the identification of potentially unfair clauses in contracts becomes crucial. In this work, we conduct an empirical study to analyze the perspectives of different stakeholders regarding contractual fairness. We then investigate the ability of Pre-trained Language Models (PLMs) to identify unfairness in contractual sentences by comparing chain of thought prompting and semi-supervised fine-tuning approaches. Using BERT-based fine-tuning, we achieved an accuracy of 84% on a dataset consisting of proprietary contracts. It outperformed chain of thought prompting using Vicuna-13B by a margin of 9%.

Statutory reasoning is the task of determining whether a given law – a part of a statute – applies to a given legal case. Previous work has shown that structured, logical representations of laws and cases can be leveraged to solve statutory reasoning, including on the StAtutory Reasoning Assessment dataset (SARA), but rely on costly human translation into structured representations. Here, we investigate a form of legal information extraction atop the SARA cases, illustrating how the task can be done with high performance. Further, we show how the performance of downstream symbolic reasoning directly correlates with the quality of the information extraction.

While large language models (LLMs) have demonstrated significant capabilities in text generation, their utilization in areas requiring domain-specific expertise, such as law, must be approached cautiously. This caution is warranted due to the inherent challenges associated with LLM-generated texts, including the potential presence of factual errors. Motivated by this issue, we propose Eval-RAG, a new evaluation method for LLM-generated texts. Unlike existing methods, Eval-RAG evaluates the validity of generated texts based on the related document that are collected by the retriever. In other words, Eval-RAG adopts the idea of retrieval augmented generation (RAG) for the purpose of evaluation. Our experimental results on Korean Legal Question-Answering (QA) tasks show that conventional LLM-based evaluation methods can be better aligned with Lawyers’ evaluations, by combining with Eval-RAG. In addition, our qualitative analysis show that Eval-RAG successfully finds the factual errors in LLM-generated texts, while existing evaluation methods cannot.

In the rapidly evolving field of crypto assets, white papers are essential documents for investor guidance, and are now subject to unprecedented content requirements under the European Union’s Markets in Crypto-Assets Regulation (MiCAR). Natural Language Processing (NLP) can serve as a powerful tool for both analyzing these documents and assisting in regulatory compliance. This paper delivers two contributions to the topic. First, we survey existing applications of textual analysis to unregulated crypto asset white papers, uncovering a research gap that could be bridged with interdisciplinary collaboration. We then conduct an analysis of the changes introduced by MiCAR, highlighting the opportunities and challenges of integrating NLP within the new regulatory framework. The findings set the stage for further research, with the potential to benefit regulators, crypto asset issuers, and investors.

In law, it is important to distinguish between obligations, permissions, prohibitions, rights, and powers. These categories are called deontic modalities. This paper evaluates the performance of two deontic modality classification models, LEGAL-BERT and a Fusion model, in a low-resource setting. To create a generalized dataset for multi-class classification, we extracted random provisions from European Union (EU) legislation. By fine-tuning previously researched and published models, we evaluate their performance on our dataset against fusion models designed for low-resource text classification. We incorporate focal loss as an alternative for cross-entropy to tackle issues of class imbalance. The experiments indicate that the fusion model performs better for both balanced and imbalanced data with a macro F1-score of 0.61 for imbalanced data, 0.62 for balanced data, and 0.55 with focal loss for imbalanced data. When focusing on accuracy, our experiments indicate that the fusion model performs better with scores of 0.91 for imbalanced data, 0.78 for balanced data, and 0.90 for imbalanced data with focal loss.

Releasing court decisions to the public relies on proper anonymization to protect all involved parties, where necessary. The Swiss Federal Supreme Court relies on an existing system that combines different traditional computational methods with human experts. In this work, we enhance the existing anonymization software using a large dataset annotated with entities to be anonymized. We compared BERT-based models with models pre-trained on in-domain data. Our results show that using in-domain data to pre-train the models further improves the F1-score by more than 5% compared to existing models. Our work demonstrates that combining existing anonymization methods, such as regular expressions, with machine learning can further reduce manual labor and enhance automatic suggestions.

The rapid advancement of Generative Artificial Intelligence (AI), particularly Large Language Models (LLMs), has led to their widespread adoption for various natural language processing (NLP) tasks. One crucial domain ripe for innovation is the Technology-Assisted Review (TAR) process in Electronic discovery (eDiscovery). Traditionally, TAR involves manual review and classification of documents for relevance over large document collections for litigations and investigations. This process is aided by machine learning and NLP tools which require extensive training and fine-tuning. In this paper, we explore the application of LLMs to TAR, specifically for predictive coding. We experiment with out-of-the-box prompting and fine-tuning of LLMs using parameter-efficient techniques. We conduct experiments using open LLMs and compare them to commercially-licensed ones. Our experiments demonstrate that open LLMs lag behind commercially-licensed models in relevance classification using out-of-the-box prompting. However, topic-specific instruction tuning of open LLMs not only improve their effectiveness but can often outperform their commercially-licensed counterparts in performance evaluations. Additionally, we conduct a user study to gauge the preferences of our eDiscovery Subject Matter Specialists (SMS) regarding human-authored versus model-generated reasoning. We demonstrate that instruction-tuned open LLMs can generate high quality reasonings that are comparable to commercial LLMs.

Applying existing question answering (QA) systems to specialized domains like law and finance presents challenges that necessitate domain expertise. Although large language models (LLMs) have shown impressive language comprehension and in-context learning capabilities, their inability to handle very long inputs/contexts is well known. Tasks specific to these domains need significant background knowledge, leading to contexts that can often exceed the maximum length that existing LLMs can process. This study explores leveraging the semi-structured nature of legal and financial data to efficiently retrieve relevant context, enabling the use of LLMs for domain-specialized QA. The resulting system outperforms contemporary models and also provides useful explanations for the answers, encouraging the integration of LLMs into legal and financial NLP systems for future research.

In legal text processing and reasoning, one normally performs information retrieval to find relevant documents of an input question, and then performs textual entailment to answer the question. The former is about relevancy whereas the latter is about affirmation (or conclusion). While relevancy and affirmation are two different concepts, there is obviously a connection between them. That is why performing retrieval and textual entailment sequentially and independently may not make the most of this mutually supportive relationship. This paper, therefore, propose a multi–task learning model for these two tasks to improve their performance. Technically, in the COLIEE dataset, we use the information of Task 4 (conclusions) to improve the performance of Task 3 (searching for legal provisions related to the question). Our empirical findings indicate that this supportive relationship truly exists. This important insight sheds light on how leveraging relationship between tasks can significantly enhance the effectiveness of our multi-task learning approach for legal text processing.

Given the complexity of the judiciary in the US Supreme Court, various procedures, along with various resources, contribute to the court system. However, most research focuses on a limited set of resources, e.g., court opinions or oral arguments, for analyzing a specific perspective in court, e.g., partisanship or voting. To gain a fuller understanding of these perspectives in the legal system of the US Supreme Court, a more comprehensive dataset, connecting different sources in different phases of the court procedure, is needed. To address this gap, we present a multi-sourced dataset for the Supreme Court, comprising court resources from different procedural phases, connecting language documents with extensive metadata. We showcase its utility through a case study on how different court documents reveal the decision direction (conservative vs. liberal) of the cases. We analyze performance differences across three protected attributes, indicating that different court resources encode different biases, and reinforcing that considering various resources provides a fuller picture of the court procedures. We further discuss how our dataset can contribute to future research directions.

This study investigates whether a legal natural language inference (NLI) model trained on the data from one US state can be transferred to another state. We fine-tuned a pre-trained model on the task of evaluating the validity of legal will statements, once with the dataset containing the Tennessee wills and once with the dataset containing the Idaho wills. Each model’s performance on the in-domain setting and the out-of-domain setting are compared to see if the models can across the states. We found that the model trained on one US state can be mostly transferred to another state. However, it is clear that the model’s performance drops in the out-of-domain setting. The F1 scores of the Tennessee model and the Idaho model are 96.41 and 92.03 when predicting the data from the same state, but they drop to 66.32 and 81.60 when predicting the data from another state. Subsequent error analysis revealed that there are two major sources of errors. First, the model fails to recognize equivalent laws across states when there are stylistic differences between laws. Second, difference in statutory section numbering system between the states makes it difficult for the model to locate laws relevant to the cases being predicted on. This analysis provides insights on how the future NLI system can be improved. Also, our findings offer empirical support to legal experts advocating the standardization of legal documents.

Realizing the recent advances from Natural Language Processing (NLP) to the legal sector poses challenging problems such as extremely long sequence lengths, specialized vocabulary that is usually only understood by legal professionals, and high amounts of data imbalance. The recent surge of Large Language Models (LLM) has begun to provide new opportunities to apply NLP in the legal domain due to their ability to handle lengthy, complex sequences. Moreover, the emergence of domain-specific LLMs has displayed extremely promising results on various tasks. In this study, we aim to quantify how general LLMs perform in comparison to legal-domain models (be it an LLM or otherwise). Specifically, we compare the zero-shot performance of three general-purpose LLMs (ChatGPT-3.5, LLaMA-70b and Falcon-180b) on the LEDGAR subset of the LexGLUE benchmark for contract provision classification. Although the LLMs were not explicitly trained on legal data, we observe that they are still able to classify the theme correctly in most cases. However, we find that their mic-F1/mac-F1 performance are upto 19.2/26.8% lesser than smaller models fine-tuned on the legal domain, thus underscoring the need for more powerful legal-domain LLMs.

Due to the ever-increasing complexity of income tax laws in the United States, the number of US taxpayers filing their taxes using tax preparation software henceforth, tax software) continues to increase. According to the U.S. Internal Revenue Service (IRS), in FY22, nearly 50% of taxpayers filed their individual income taxes using tax software. Given the legal consequences of incorrectly filing taxes for the taxpayer, ensuring the correctness of tax software is of paramount importance. Metamorphic testing has emerged as a leading solution to test and debug legal-critical tax software due to the absence of correctness requirements and trustworthy datasets. The key idea behind metamorphic testing is to express the properties of a system in terms of the relationship between one input and its slightly metamorphosed twinned input. Extracting metamorphic properties from IRS tax publications is a tedious and time-consuming process. As a response, this paper formulates the task of generating metamorphic specifications as a translation task between properties extracted from tax documents - expressed in natural language - to a contrastive first-order logic form. We perform a systematic analysis on the potential and limitations of in-context learning with Large Language Models (LLMs) for this task, and outline a research agenda towards automating the generation of metamorphic specifications for tax preparation software.

Advancements in natural language processing (NLP) and language models have demonstrated immense potential in the legal domain, enabling automated analysis and comprehension of legal texts. However, developing robust models in Legal NLP is significantly challenged by the scarcity of resources. This paper presents AsyLex, the first dataset specifically designed for Refugee Law applications to address this gap. The dataset introduces 59,112 documents on refugee status determination in Canada from 1996 to 2022, providing researchers and practitioners with essential material for training and evaluating NLP models for legal research and case review. Case review is defined as entity extraction and outcome prediction tasks. The dataset includes 19,115 gold-standard human-labeled annotations for 20 legally relevant entity types curated with the help of legal experts and 1,682 gold-standard labeled documents for the case outcome. Furthermore, we supply the corresponding trained entity extraction models and the resulting labeled entities generated through the inference process on AsyLex. Four supplementary features are obtained through rule-based extraction. We demonstrate the usefulness of our dataset on the legal judgment prediction task to predict the binary outcome and test a set of baselines using the text of the documents and our annotations. We observe that models pretrained on similar legal documents reach better scores, suggesting that acquiring more datasets for specialized domains such as law is crucial.

In this study, we explore the performance oflarge language models (LLMs) using differ-ent prompt engineering approaches in the con-text of legal text classification. Prior researchhas demonstrated that various prompting tech-niques can improve the performance of a di-verse array of tasks done by LLMs. However,in this research, we observe that professionaldocuments, and in particular legal documents,pose unique challenges for LLMs. We experi-ment with several LLMs and various promptingtechniques, including zero/few-shot prompting,prompt ensembling, chain-of-thought, and ac-tivation fine-tuning and compare the perfor-mance on legal datasets. Although the newgeneration of LLMs and prompt optimizationtechniques have been shown to improve gener-ation and understanding of generic tasks, ourfindings suggest that such improvements maynot readily transfer to other domains. Specifi-cally, experiments indicate that not all prompt-ing approaches and models are well-suited forthe legal domain which involves complexitiessuch as long documents and domain-specificlanguage.

U.S. Federal Regulators receive over one million comment letters each year from businesses, interest groups, and members of the public, all advocating for changes to proposed regulations. These comments are believed to have wide-ranging impacts on public policy. However, measuring the impact of specific comments is challenging because regulators are required to respond to comments but they do not have to specify which comments they are addressing. In this paper, we propose a simple yet effective solution to this problem by using an iterative contrastive method to train a neural model aiming for matching text from public comments to responses written by regulators. We demonstrate that our proposal substantially outperforms a set of selected text-matching baselines on a human-annotated test set. Furthermore, it delivers performance comparable to the most advanced gigantic language model (i.e., GPT-4), and is more cost-effective when handling comments and regulator responses matching in larger scale.

With 102,530,067 items currently in its crowd-sourced knowledge base, Wikidata provides NLP practitioners a unique and powerful resource for inference and reasoning over real-world entities. However, because Wikidata is very entity focused, events and actions are often labeled with eventive nouns (e.g., the process of diagnosing a person’s illness is labeled “diagnosis”), and the typical participants in an event are not described or linked to that event concept (e.g., the medical professional or patient). Motivated by a need for an adaptable, comprehensive, domain-flexible ontology for information extraction, including identifying the roles entities are playing in an event, we present a curated subset of Wikidata in which events have been enriched with PropBank roles. To enable richer narrative understanding between events from Wikidata concepts, we have also provided a comprehensive mapping from temporal Qnodes and Pnodes to the Allen Interval Temporal Logic relations.

The paper presents the work on the selection, semantic annotation and classification of a group of verbs from WordNet, characterized with the semantic primitive ‘verbs of contact’ that belong to the common Bulgarian lexis. The selection of the verb set using both different criteria: statistical information from corpora, WordNet Base concepts and AoA as a criterion, is described. The focus of the work is on the process of the verbs’ of contact semantic annotation using the combined information from two language resources - WordNet and FrameNet. The verbs of contact from WordNet are assigmed semantic frames from FrameNet and then grouped in semantic subclasses using both their place in the WordNet hierarchy, the semantic restrictions on their frame elements and the corresponding syntactic realization. At the end we offer some conclusions on the classification of ‘verbs of contact’ in semantic subtypes.

In this work, we address the annotation of language resources through the application of the engagement network in appraisal theory. This work represents an attempt to extend the advances in studies of speech and dialogue acts to encompass the latest notion of stance negotiations in discourse, between the writer and other sources. This type of phenomenon has become especially salient in contemporary media communication and requires some timely research to address emergent requirement. We shall first of all describe the engagement network as proposed by Martin and White (2005) and then discuss the issue of multisubjectivity. We shall then propose and describe a bi-step procedure towards better annotation before discussing the benefits of engagement network in the assessment of speaker-writer stance. We shall finally discuss issues of annotation consistency and reliability.

This paper presents a graphical editor for directed graphs, serialised in the PENMAN format, as used for annotations in Abstract Meaning Representation (AMR). The tool supports creating and modifying of AMR graphs and other directed graphs, adding and deletion of instances, edges and literals, renaming of concepts, relations and literals, setting a “top node” and validating the edited graph.

Personal nouns, i.e. common nouns denoting human beings, play an important role in manifesting gender and gender stereotypes in texts, especially for languages with grammatical gender like German. Automatically detecting and extracting personal nouns can thus be of interest to a myriad of different tasks such as minimizing gender bias in language models and researching gender stereotypes or gender-fair language, but is complicated by the morphological heterogeneity and homonymy of personal and non-personal nouns, which restrict lexicon-based approaches. In this paper, we introduce a classifier created by fine-tuning a transformer model that detects personal nouns in German. Although some phenomena like homonymy and metalinguistic uses are still problematic, the model is able to classify personal nouns with robust accuracy (f1-score: 0.94).

The article discusses the challenges of cross-linguistic dialogue act annotation, which involves using methods developed for one language to annotate conversations in another language. The article specifically focuses on the research on dialogue act annotation in Polish, based on the ISO standard developed for English. The article examines the differences between Polish and English in dialogue act annotation based on selected examples from DiaBiz.Kom corpus, such as the use of honorifics in Polish, the use of inflection to convey meaning in Polish, the tendency to use complex sentence structures in Polish, and the cultural differences that may play a role in the annotation of dialogue acts. The article also discusses the creation of DiaBiz.Kom, a Polish dialogue corpus based on ISO 24617-2 standard applied to 1100 transcripts.

Using recent developments in count noun quantification, namely Referential Transparency Theory (RTT), the basic structure for annotating quantification in the nominal domain according to RTT is presented. The paper discusses core ideas of RTT, derives the abstract annotation syntax, and exemplifies annotations of quantified noun phrases partly in comparison to QuantML.

This paper discusses some issues in the semantic annotation of quantification phenomena in general, and in particular in the markup language QuantML, which has been proposed to form part of an ISO standard annotation scheme for quantification in natural language data. QuantML annotations have been claimed to have a compositional semantic interpretation, but the formal specification of QuantML in the official ISO documentation does not provide sufficient detail to judge this. This paper aims to fill this gap.

VoxML is a modeling language used to map natural language expressions into real time visualizations using real-world semantic knowledge of objects and events. Its utility has been demonstrated in embodied simulation environmens and in agent-object interactions in situated human-agent communicative. It is enriched to work with notions of affordances, both Gibsonian and Telic, and habitat for various interactions between the rational agent (human) and an object. This paper aims to specify VoxML as an annotation language in general abstract terms. It then shows how it works on annotating linguistic data that express visually perceptible human-object interactions. The annotation structures thus generated will be interpreted against the enriched minimal model created by VoxML as a modeling language while supporting the modeling purposes of VoxML linguistically.

In this work, we assess the quality of different utterance segmentation techniques as an aid in annotating collaborative problem solving in teams and the creation of shared meaning between participants in a situated, collaborative task. We manually transcribe utterances in a dataset of triads collaboratively solving a problem involving dialogue and physical object manipulation, annotate collaborative moves according to these gold-standard transcripts, and then apply these annotations to utterances that have been automatically segmented using toolkits from Google and Open-AI’s Whisper. We show that the oracle utterances have minimal correspondence to automatically segmented speech, and that automatically segmented speech using different segmentation methods is also inconsistent. We also show that annotating automatically segmented speech has distinct implications compared with annotating oracle utterances — since most annotation schemes are designed for oracle cases, when annotating automatically-segmented utterances, annotators must make arbitrary judgements which other annotators may not replicate. We conclude with a discussion of how future annotation specs can account for these needs.

The area of designing semantic/meaning representations is a dynamic one with new formalisms and extensions being proposed continuously. It may be challenging for users of semantic representations to select the relevant formalism for their purpose or for newcomers to the field to select the features they want to represent in a new formalism. In this paper, we propose a set of structural and global features to consider when designing formalisms, and against which formalisms can be compared. We also propose a sample comparison of a number of existing formalisms across the selected features, complemented by a more entailment-oriented comparison on the phenomena of the FraCaS corpus.

Universal Semantic Representation (USR) is designed as a language-independent information packaging system that captures information at three levels: (a) Lexico-conceptual, (b) Syntactico-Semantic, and (c) Discourse. Unlike other representations that mainly encode predicates and their argument structures, our proposed representation captures the speaker’s vivakṣā- how the speaker views the activity. The idea of “speaker’s vivakṣā is inspired by Indian Grammatical Tradition. There can be some amount of idiosyncrasy of the speaker in the annotation since it is the speaker’s view- point that has been captured in the annotation. Hence the evaluation metrics of such resources need to be also thought through from scratch. This paper presents an extensive evaluation procedure of this semantic representation from two perspectives (a) Inter- Annotator Agreement and (b) one downstream task, namely multilingual Natural Language Generation. We also qualitatively evaluate the experience of natural language generation by manual parsing of USR, so as to understand the readability of USR. We have achieved above 80% Inter-Annotator Agreement for USR annotations and above 80% semantic closeness in multi-lingual generation tasks suggesting the reliability of USR annotations and utility for multi-lingual generations. The qualitative evaluation also suggests high readability and hence the utility of USR as a semantic representation.

Abstract Meaning Representation (AMR) is a popular semantic annotation schema that presents sentence meaning as a graph while abstracting away from syntax. It was originally designed for English, but has since been extended to a variety of non-English versions of AMR. These cross-lingual adaptations, to varying degrees, incorporate language-specific features necessary to effectively capture the semantics of the language being annotated. Uniform Meaning Representation (UMR) on the other hand, the multilingual extension of AMR, was designed specifically for cross-lingual applications. In this work, we discuss these two approaches to extending AMR beyond English. We describe both approaches, compare the information they capture for a case language (Spanish), and outline implications for future work.

To collaborate effectively in physically situated tasks, robots must be able to ground concepts in natural language to the physical objects in the environment as well as their own capabilities. We describe the implementation and the demonstration of a system architecture that sup- ports tasking robots using natural language. In this architecture, natural language instructions are first handled by a dialogue management component, which provides feedback to the user and passes executable instructions along to an Abstract Meaning Representation (AMR) parser. The parse distills the action primitives and parameters of the instructed behavior in the form of a directed a-cyclic graph, passed on to the grounding component. We find AMR to be an efficient formalism for grounding the nodes of the graph using a Distributed Correspondence Graph. Thus, in our approach, the concepts of language are grounded to entities in the robot’s world model, which is populated by its sensors, thereby enabling grounded natural language communication. The demonstration of this system will allow users to issue navigation commands in natural language to direct a simulated ground robot (running the Robot Operating System) to various landmarks observed by the user within a simulated environment.

Actions are critical for interpreting dialogue: they provide context for demonstratives and definite descriptions in discourse, and they continually update the common ground. This paper describes how Abstract Meaning Representation (AMR) can be used to annotate actions in multimodal human-human and human-object interactions. We conduct initial annotations of shared task and first-person point-of-view videos. We show that AMRs can be interpreted by a proxy language, such as VoxML, as executable annotation structures in order to recreate and simulate a series of annotated events.

Accurately interpreting the relationships between actions in a recipe text is essential to successful recipe completion. We explore using Abstract Meaning Representation (AMR) to represent recipe instructions, abstracting away from syntax and sentence structure that may order recipe actions in arbitrary ways. We present an algorithm to split sentence-level AMRs into action-level AMRs for individual cooking steps. Our approach provides an automatic way to derive fine-grained AMR representations of actions in cooking recipes and can be a useful tool for downstream, instructional tasks.

This work compares two ways of annotating semantic relations expressed in prepositional phrases: semantic classes in the Semantic Network of Adposition and Case Supersenses (SNACS), and tectogrammatical functors from the Prague English Dependency Treebank (PEDT). We compare the label definitions in the respective annotation guidelines to determine expected mappings, then check how well these work empirically using Wall Street Journal text. In the definitions we find substantial overlap in the distributions of the two schemata with respect to participants and circumstantials, but substantial divergence for configurational relationships between nominals. This is borne out by the empirical analysis. Examining the data more closely for participants and circumstantials reveals that there are some unexpected, yet systematic divergences between definitionally aligned groups.

Identifying all predicate-argument relations in a sentence has been a fundamental research target in NLP. While traditionally these relations were modeled via formal schemata, the recent QA-SRL paradigm (and its extensions) present appealing advantages of capturing such relations through intuitive natural language question-answer (QA) pairs. In this paper, we extend the QA-based semantics framework to cover adjectival predicates, which carry important information in many downstream settings yet have been scarcely addressed in NLP research. Firstly, based on some prior literature and empirical assessment, we propose capturing four types of core adjectival arguments, through corresponding question types. Notably, our coverage goes beyond prior annotations of adjectival arguments, while also explicating valuable implicit arguments. Next, we develop an extensive data annotation methodology, involving controlled crowdsourcing and targeted expert review. Following, we create a high-quality dataset, consisting of 9K adjective mentions with 12K predicate-argument instances (QAs). Finally, we present and analyze baseline models based on text-to-text language modeling, indicating challenges for future research, particularly regarding the scarce argument types. Overall, we suggest that our contributions can provide the basis for research on contemporary modeling of adjectival information.

This paper presents a new dataset with Discourse Representation Structures (DRSs) annotated over naturally-occurring sentences. Importantly, these sentences are more varied in length and on average longer than those in the existing gold-standard DRS dataset, the Parallel Meaning Bank, and we show that they are therefore much harder for parsers. We argue, though, that this provides a more realistic assessment of the difficulties of DRS parsing.

Rooted in AMR, Uniform Meaning Representation (UMR) is a graph-based formalism with nodes as concepts and edges as relations between them. When used to represent natural language semantics, UMR maps words in a sentence to concepts in the UMR graph. Multiword expressions (MWEs) pose a particular challenge to UMR annotation because they deviate from the default one-to-one mapping between words and concepts. There are different types of MWEs which require different kinds of annotation that must be specified in guidelines. This paper discusses the specific treatment for each type of MWE in UMR.

Despite the significant progress made in Natural Language Processing (NLP) thanks to deep learning techniques, efforts are still needed to model explicit, factual, and accurate meaning representation formalisms. In this article, we present a comparative table of ten formalisms that have been proposed over the last thirty years, and we describe and put forth our own, Meaning Representation for Application Purposes (MR4AP), developed in an industrial context with a definitive applicative aim.

We propose the use of modal dependency parses (MDPs) aligned with syntactic dependency parse trees as an avenue for the novel task of claim extraction. MDPs provide a document-level structure that links linguistic expression of events to the conceivers responsible for those expressions. By defining the event-conceiver links as claims and using subgraph pattern matching to exploit the complementarity of these modal links and syntactic claim patterns, we outline a method for aggregating and classifying claims, with the potential for supplying a novel perspective on large natural language data sets. Abstracting away from the task of claim extraction, we prototype an interpretable information extraction (IE) paradigm over sentence- and document-level parse structures, framing inference as subgraph matching and learning as subgraph mining. We make our code open-sourced at https://github.com/BBN-E/nlp-graph-pattern-matching-and-mining.

The expansion of Large Language Models (LLMs) into more serious areas of application, involving decision-making and the forming of public opinion, calls for a more thoughtful treatment of texts. Augmenting them with explicit and understandable argumentative analysis could foster a more reasoned usage of chatbots, text completion mechanisms or other applications. However, it is unclear which aspects of argumentation can be reliably identified and integrated by them. In this paper we propose an adaptation of Wagemans (2016)’s Periodic Table of Arguments to identify different argumentative aspects of texts, with a special focus on hate speech in social media. We have empirically assessed the reliability with which each of these aspects can be automatically identified. We analyze the implications of these results, and how to adapt the proposal to obtain reliable representations of those that cannot be successfully identified.

We evaluate the performance of various models on the task of named entity recognition (NER) for classical Latin. Using an existing dataset, we train two transformer-based LatinBERT models and one shallow conditional random field (CRF) model. The performance is assessed using both standard metrics and a detailed manual error analysis, and compared to the results obtained by different already released Latin NER tools. Both analyses demonstrate that the BERT models achieve a better f1-score than the other models. Furthermore, we annotate new, unseen data for further evaluation of the models, and we discuss the impact of annotation choices on the results.

Contextual language models have been trained on Classical languages, including Ancient Greek and Latin, for tasks such as lemmatization, morphological tagging, part of speech tagging, authorship attribution, and detection of scribal errors. However, high-quality sentence embedding models for these historical languages are significantly more difficult to achieve due to the lack of training data. In this work, we use a multilingual knowledge distillation approach to train BERT models to produce sentence embeddings for Ancient Greek text. The state-of-the-art sentence embedding approaches for high-resource languages use massive datasets, but our distillation approach allows our Ancient Greek models to inherit the properties of these models while using a relatively small amount of translated sentence data. We build a parallel sentence dataset using a sentence-embedding alignment method to align Ancient Greek documents with English translations, and use this dataset to train our models. We evaluate our models on translation search, semantic similarity, and semantic retrieval tasks and investigate translation bias. We make our training and evaluation datasets freely available.

In this project we train a Transformer-based model from scratch, with the goal of parsing the morphology of Ancient Syriac texts as accurately as possible. Syriac is still a low resource language, only a relatively small training set was available. Therefore, the training set was expanded by adding Biblical Hebrew data to it. Five different experiments were done: the model was trained on Syriac data only, it was trained with mixed Syriac and (un)vocalized Hebrew data, and it was pretrained on (un)vocalized Hebrew data and then finetuned on Syriac data. The models trained on Hebrew and Syriac data consistently outperform the models trained on Syriac data only. This shows, that the differences between Syriac and Hebrew are small enough that it is worth adding Hebrew data to train the model for parsing Syriac morphology. Training models on different languages is an important trend in NLP, we show that this works well for relatively small datasets of Syriac and Hebrew.

Intertextual allusions hold a pivotal role in Classical Philology, with Latin authors frequently referencing Ancient Greek texts. Until now, the automatic identification of these intertextual references has been constrained to monolingual approaches, seeking parallels solely within Latin or Greek texts. In this study, we introduce SPhilBERTa, a trilingual Sentence-RoBERTa model tailored for Classical Philology, which excels at cross-lingual semantic comprehension and identification of identical sentences across Ancient Greek, Latin, and English. We generate new training data by automatically translating English into Ancient Greek texts. Further, we present a case study, demonstrating SPhilBERTa’s capability to facilitate automated detection of intertextual parallels. Intertextual allusions hold a pivotal role in Classical Philology, with Latin authors frequently referencing Ancient Greek texts. Until now, the automatic identification of these intertextual references has been constrained to monolingual approaches, seeking parallels solely within Latin or Greek texts. In this study, we introduce SPhilBERTa, a trilingual Sentence-RoBERTa model tailored for Classical Philology, which excels at cross-lingual semantic comprehension and identification of identical sentences across Ancient Greek, Latin, and English. We generate new training data by automatically translating English into Ancient Greek texts. Further, we present a case study, demonstrating SPhilBERTa’s capability to facilitate automated detection of intertextual parallels.

Etruscan is an ancient language spoken in Italy from the 7th century BC to the 1st century AD. There are no native speakers of the language at the present day, and its resources are scarce, as there are an estimated 12,000 known inscriptions. To the best of our knowledge, there are no publicly available Etruscan corpora for natural language processing. Therefore, we propose a dataset for machine translation from Etruscan to English, which contains 2891 translated examples from existing academic sources. Some examples are extracted manually, while others are acquired in an automatic way. Along with the dataset, we benchmark different machine translation models observing that it is possible to achieve a BLEU score of 10.1 with a small transformer model. Releasing the dataset can help enable future research on this language, similar languages or other languages with scarce resources.

We evaluate four count-based and predictive distributional semantic models of Ancient Greek against AGREE, a composite benchmark of human judgements, to assess their ability to retrieve semantic relatedness. On the basis of the observations deriving from the analysis of the results, we design a procedure for a larger-scale intrinsic evaluation of count-based and predictive language models, including syntactic embeddings. We also propose possible ways of exploiting the different layers of the whole AGREE benchmark (including both human- and machine-generated data) and different evaluation metrics.

This paper focuses on the process of harmonising the five Latin treebanks available in Universal Dependencies with respect to morphological annotation. We propose a workflow that allows to first spot inconsistencies and missing information, in order to detect to what extent the annotations differ, and then correct the retrieved bugs, with the goal of equalising the annotation of morphological features in the treebanks and producing more consistent linguistic data. Subsequently, we present some experiments carried out with UDPipe and Stanza in order to assess the impact of such harmonisation on parsing accuracy.

Constituency parsing plays a fundamental role in advancing natural language processing (NLP) tasks. However, training an automatic syntactic analysis system for ancient languages solely relying on annotated parse data is a formidable task due to the inherent challenges in building treebanks for such languages. It demands extensive linguistic expertise, leading to a scarcity of available resources. To overcome this hurdle, cross-lingual transfer techniques which require minimal or even no annotated data for low-resource target languages offer a promising solution. In this study, we focus on building a constituency parser for Middle High German (MHG) under realistic conditions, where no annotated MHG treebank is available for training. In our approach, we leverage the linguistic continuity and structural similarity between MHG and Modern German (MG), along with the abundance of MG treebank resources. Specifically, by employing the delexicalization method, we train a constituency parser on MG parse datasets and perform cross-lingual transfer to MHG parsing. Our delexicalized constituency parser demonstrates remarkable performance on the MHG test set, achieving an F1-score of 67.3%. It outperforms the best zero-shot cross-lingual baseline by a margin of 28.6% points. The encouraging results underscore the practicality and potential for automatic syntactic analysis in other ancient languages that face similar challenges as MHG.

Large language models (LLMs) have demonstrated exceptional language understanding and generation capabilities. However, their ability to comprehend ancient languages, specifically ancient Chinese, remains largely unexplored. To bridge this gap, we introduce ACLUE, an evaluation benchmark designed to assess the language abilities of models in relation to ancient Chinese. ACLUE consists of 15 tasks that cover a range of skills, including phonetic, lexical, syntactic, semantic, inference and knowledge. By evaluating 8 state-of-the-art multilingual and Chinese LLMs, we have observed a significant divergence in their performance between modern Chinese and ancient Chinese. Among the evaluated models, ChatGLM2 demonstrates the highest level of performance, achieving an average accuracy of 37.45%. We have established a leaderboard for communities to assess their models.

This ongoing study explores emotion recognition in Latin texts, specifically focusing on Latin comedies. Leveraging Natural Language Processing and classical philology insights, the project navigates the challenges of Latin’s intricate grammar and nuanced emotional expression. Despite initial challenges with lexicon translation and emotional alignment, the work provides a foundation for a more comprehensive analysis of emotions in Latin literature.

Machine translation (MT) of ancient Chinese texts presents unique challenges due to the complex grammatical structures, cultural nuances, and polysemy of the language. This paper focuses on evaluating the translation quality of different platforms for ancient Chinese texts using The Analects as a case study. The evaluation is conducted using the BLEU, LMS, and ESS metrics, and the platforms compared include three machine translation platforms (Baidu Translate, Bing Microsoft Translator, and DeepL), and one language generation model ChatGPT that can engage in translation endeavors. Results show that Baidu performs the best, surpassing the other platforms in all three metrics, while ChatGPT ranks second and demonstrates unique advantages. The translations generated by ChatGPT are deemed highly valuable as references. The study contributes to understanding the challenges of MT for ancient Chinese texts and provides insights for users and researchers in this field. It also highlights the importance of considering specific domain requirements when evaluating MT systems.

The Bavarian Academy of Sciences and Humanities aims to digitize the Medieval Latin Dictionary. This dictionary entails record cards referring to lemmas in medieval Latin, a low-resource language. A crucial step of the digitization process is the handwritten text recognition (HTR) of the handwritten lemmas on the record cards. In our work, we introduce an end-to-end pipeline, tailored for the medieval Latin dictionary, for locating, extracting, and transcribing the lemmas. We employ two state-of-the-art image segmentation models to prepare the initial data set for the HTR task. Further, we experiment with different transformer-based models and conduct a set of experiments to explore the capabilities of different combinations of vision encoders with a GPT-2 decoder. Additionally, we also apply extensive data augmentation resulting in a highly competitive model. The best-performing setup achieved a character error rate of 0.015, which is even superior to the commercial Google Cloud Vision model, and shows more stable performance.

This paper reports on the results of a comparative evaluation in view of the development of a new lemmatizer for unedited, Byzantine Greek texts. For the experiment, the performance of four existing lemmatizers, all pre-trained on Ancient Greek texts, was evaluated on how well they could handle texts stemming from the Middle Ages and displaying quite some peculiarities. The aim of this study is to get insights into the pitfalls of existing lemmatistion approaches as well as the specific challenges of our Byzantine Greek corpus, in order to develop a lemmatizer that can cope with its peculiarities. The results of the experiment show an accuracy drop of 20pp. on our corpus, which is further investigated in a qualitative error analysis.

Commentary of Gongyang, Commentary of Guliang, and Commentary of Zuo are collectively called the Three Commentaries on the Spring and Autumn Annals, which are the supplement and interpretation of the content of Spring and Autumn Annals with value in historical and literary research. In traditional research paradigms, scholars often explored the differences between the Three Commentaries within the details in contexts. Starting from the view of computational humanities, this paper examines the differences in the language style of the Three Commentaries through the representation of language, which takes the methods of deep learning. Specifically, this study vectorizes the context at word and sentence levels. It maps them into the same plane to find the differences between the use of words and sentences in the Three Commentaries. The results show that the Commentary of Gongyang and the Commentary of Guliang are relatively similar, while the Commentary of Zuo is significantly different. This paper verifies the feasibility of deep learning methods in stylistics study under computational humanities. It provides a valuable perspective for studying the Three Commentaries on the Spring and Autumn Annals.

The digitization of ancient books necessitates the implementation of automatic word segmentation and part-of-speech tagging. However, the existing research on this topic encounters pressing issues, including suboptimal efficiency and precision, which require immediate resolution. This study employs a methodology that combines word segmentation and part-of-speech tagging. It establishes a correlation between fonts and radicals, trains the Radical2Vec radical vector representation model, and integrates it with the SikuRoBERTa word vector representation model. Finally, it connects the BiLSTM-CRF neural network.The study investigates the combination of word segmentation and part-of-speech tagging through an experimental approach using a specific data set. In the evaluation dataset, the F1 score for word segmentation is 95.75%, indicating a high level of accuracy. Similarly, the F1 score for part-of-speech tagging is 91.65%, suggesting a satisfactory performance in this task. This model enhances the efficiency and precision of the processing of ancient books, thereby facilitating the advancement of digitization efforts for ancient books and ensuring the preservation and advancement of ancient book heritage.

The study of Sumerian texts often requires domain experts to examine a vast number of tables. However, the absence of user-friendly tools for this process poses challenges and consumes significant time. In addressing this issue, we introduce an open-source library that empowers domain experts with minimal technical expertise to automate manual and repetitive tasks using a no-code dashboard. Our library includes an information extraction module that enables the automatic extraction of names and relations based on the user-defined lists of name tags and relation types. By utilizing the tool to facilitate the creation of knowledge graphs which is a data representation method offering insights into the relationships among entities in the data, we demonstrate its practical application in the analysis of Sumerian texts.

Abstract : Based on the oracle bone glyph data in the “ZhongHuaZiKu”database, this paper designs a new input method coding scheme which is easy to search in the database, and provides a feasible scheme for the design of oracle bone glyph input method software in the future. The coding scheme in this paper is based on the experience of the past oracle bone inscriptions input method design. In view of the particularity of oracle bone inscriptions, the difference factors such as component combination, sound code and shape code ( letter ) are added, and the coding format is designed as follows : The single component characters in the identified characters are arranged according to the format of " structural code + pronunciation full spelling code + tone code " ; the multi-component characters in the identified characters are arranged according to the format of " structure code + split component pronunciation full spelling code + overall glyph pronunciation full spelling code”; unidentified characters are arranged according to the format of " y + identified component pronunciation full spelling + unidentified component shape code ( letter ) ".Among them, the identified component code and the unidentified component shape code are input in turn according to the specific glyph from left to right, from top to bottom, and from outside to inside. Encoding through these coding formats, the heavy code rate is low, and the input habits of most people are also taken into account. Keywords : oracle bone inscriptions ; input method ; coding

This paper seeks to leverage translations of Ancient Greek texts to enhance the performance of automatic word sense disambiguation (WSD). Satisfactory WSD in Ancient Greek is achievable, provided that the system can rely on annotated data. This study, acknowledging the challenges of manually assigning meanings to every Greek lemma, explores the strategies to derive WSD data from parallel texts using sentence and word alignment. Our results suggest that, assuming the condition of high word frequency is met, this technique permits us to automatically produce a significant volume of annotated data, although there are still significant obstacles when trying to automate this process.

Classical Arabic, like all other historical languages, lacks adequate training datasets and accurate “off-the-shelf” models that can be directly employed in the processing pipelines. In this paper, we present our in-progress work in developing and training deep learning models tailored for handling diverse tasks relevant to classical Arabic texts. Specifically, we focus on Named Entities Recognition, person relationships classification, toponym sub-classification, onomastic section boundaries detection, onomastic entities classification, as well as date recognition and classification. Our work aims to address the challenges associated with these tasks and provide effective solutions for analyzing classical Arabic texts. Although this work is still in progress, the preliminary results reported in the paper indicate excellent to satisfactory performance of the fine-tuned models, effectively meeting the intended goal for which they were trained.

We present statistical and machine-learning based techniques for detecting and correcting errors in text and apply them to the challenge of textual corruption in Greek philology. Most ancient Greek texts reach us through a long process of copying, in relay, from earlier manuscripts (now lost). In this process of textual transmission, copying errors tend to accrue. After training a BERT model on the largest premodern Greek dataset used for this purpose to date, we identify and correct previously undetected errors made by scribes in the process of textual transmission, in what is, to our knowledge, the first successful identification of such errors via machine learning. The premodern Greek BERT model we train is available for use at https://huggingface.co/cabrooks/LOGION-base.

This paper provides an overview of diverse applications of parallel corpora in ancient languages, particularly Ancient Greek. In the first part, we provide the fundamental principles of parallel corpora and a short overview of their applications in the study of ancient texts. In the second part, we illustrate how to leverage on parallel corpora to perform various NLP tasks, including automatic translation alignment, dynamic lexica induction, and Named Entity Recognition. In the conclusions, we emphasize current limitations and future work.

Research into emotions is a developing field within Assyriology, and NLP tools for Akkadian texts offers a new perspective on the data. In this submission, we use PMI-based word embeddings to explore the relationship between parts of the body and emotions. Using data downloaded from Oracc, we ask which parts of the body were semantically linked to emotions. We do this through examining which of the top 10 results for a body part could be used to express emotions. After identifying two words for the body that have the most emotion words in their results list (libbu and kabattu), we then examine whether the emotion words in their results lists were indeed used in this manner in the Neo-Assyrian textual corpus. The results indicate that of the two body parts, kabattu was semantically linked to happiness and joy, and had a secondary emotional field of anger.

We presented a pipeline for POS-tagging and lemmatizing cuneiform languages and evaluated its performance on Sumerian, first millennium Babylonian, Neo-Assyrian and Urartian texts extracted from Oracc. The system achieves a POS-tagging accuracy between 95-98% and a lemmatization accuracy of 94-96% depending on the language or dialect. For OOV words only, the current version can predict correct POS-tags for 83-91%, and lemmata for 68-84% of the input words. Compared with the earlier version, the current one has about 10% higher accuracy in OOV lemmatization and POS-tagging due to better neural network performance. We also tested the system for lemmatizing and POS-tagging the PROIEL Ancient Greek and Latin treebanks, achieving results similar to those with the cuneiform languages.

Dependency parsing is a syntactic analysis method to analyze the dependency relationships between words in a sentence. The interconnection between words through dependency relationships is typical graph data. Traditional Tibetan dependency parsing methods typically model dependency analysis as a transition-based or sequence-labeling task, ignoring the graph information between words. To address this issue, this paper proposes a graph neural network (GNN)-based Tibetan dependency parsing method. This method treats Tibetan words as nodes and the dependency relationships between words as edges, thereby constructing the graph data of Tibetan sentences. Specifically, we use BiLSTM to learn the word representations of Tibetan, utilize GNN to model the relationships between words and employ MLP to predict the types of relationships between words. We conduct experiments on a Tibetan dependency database, and the results show that the proposed method can achieve high-quality Tibetan dependency parsing results.

Ancient ethnic documents are essential to China’s ancient literature and an indispensable civilizational achievement of Chinese culture. However, few research teams are involved due to language and script literacy limitations. To address these issues, this paper proposes an interlinearized annotation strategy for ancient ethnic literature. This strategy aims to alleviate text literacy difficulties, encourage interdisciplinary researchers to participate in studying ancient ethnic literature, and improve the efficiency of ancient ethnic literature development. Concretely, the interlinearized annotation consists of original, word segmentation, Latin, annotated, and translation lines. In this paper, we take ancient Tibetan literature as an example to explore the interlinearized annotation strategy. However, manually building large-scale corpus is challenging. To build a large-scale interlinearized dataset, we propose a multi-task learning-based interlinearized annotation method, which can generate interlinearized annotation lines based on the original line. Experimental results show that after training on about 10,000 sentences (lines) of data, our model achieves 70.9% and 63.2% F1 values on the segmentation lines and annotated lines, respectively, and 18.7% BLEU on the translation lines. It dramatically enhances the efficiency of data annotation, effectively speeds up interlinearized annotation, and reduces the workload of manual annotation.

Although image captioning has a vast array of applications, it has not reached its full potential in languages other than English. Arabic, for instance, although the native language of more than 400 million people, remains largely underrepresented in this area. This is due to the lack of labeled data and powerful Arabic generative models. We alleviate this issue by presenting a novel vision-language model dedicated to Arabic, dubbed Violet. Our model is based on a vision encoder and a Gemini text decoder that maintains generation fluency while allowing fusion between the vision and language components. To train our model, we introduce a new method for automatically acquiring data from available English datasets. We also manually prepare a new dataset for evaluation. Violet performs sizeably better than our baselines on all of our evaluation datasets. For example, it reaches a CIDEr score of 61.2 on our manually annotated dataset and achieves an improvement of 13 points on Flickr8k.

This paper presents Nâbra (نَبْرَة), a corpora of Syrian Arabic dialects with morphological annotations. A team of Syrian natives collected more than 6K sentences containing about 60K words from several sources including social media posts, scripts of movies and series, lyrics of songs and local proverbs to build Nâbra. Nâbra covers several local Syrian dialects including those of Aleppo, Damascus, Deir-ezzur, Hama, Homs, Huran, Latakia, Mardin, Raqqah, and Suwayda. A team of nine annotators annotated the 60K tokens with full morphological annotations across sentence contexts. We trained the annotators to follow methodological annotation guidelines to ensure unique morpheme annotations, and normalized the annotations. F1 and 𝜅 agreement scores ranged between 74% and 98% across features, showing the excellent quality of Nâbra annotations. Our corpora are open-source and publicly available as part of the Currasat portal https://sina.birzeit.edu/currasat.

Arabic is one of the most globally spoken languages with more than 313 million speakers worldwide. Arabic handwriting is known for its cursive nature and the variety of writing styles used. Despite the increase in effort to digitize artistic and historical elements, no public dataset was released to deal with Arabic text recognition for realistic manuscripts and calligraphic text. We present the Handwriting Identification of Manuscripts and Calligraphy in Arabic (HICMA) dataset as the first publicly available dataset with real-world and diverse samples of Arabic handwritten text in manuscripts and calligraphy. With more than 5,000 images across five different styles, the HICMA dataset includes image-text pairs and style labels for all images. We further present a comparison of the current state-of-the-art optical character recognition models in Arabic and benchmark their performance on the HICMA dataset, which serves as a baseline for future works. Both the HICMA dataset and its benchmarking tool are made available to the public under the CC BY-NC 4.0 license in the hope that the presented work opens the door to further enhancements of complex Arabic text recognition.

As Electronic Health Records (EHR) become ubiquitous in healthcare systems worldwide, including in Arabic-speaking countries, the dual imperative of safeguarding patient privacy and leveraging data for research and quality improvement grows. This paper presents a first-of-its-kind automated de-identification pipeline for medical text specifically tailored for the Arabic language. This includes accurate medical Named Entity Recognition (NER) for identifying personal information; data obfuscation models to replace sensitive entities with fake entities; and an implementation that natively scales to large datasets on commodity clusters. This research makes two contributions. First, we adapt two existing NER architectures— BERT For Token Classification (BFTC) and BiLSTM-CNN-Char – to accommodate the unique syntactic and morphological characteristics of the Arabic language. Comparative analysis suggests that BFTC models outperform Bi-LSTM models, achieving higher F1 scores for both identifying and redacting personally identifiable information (PII) from Arabic medical texts. Second, we augment the deep learning models with a contextual parser engine to handle commonly missed entities. Experiments show that the combined pipeline demonstrates superior performance with micro F1 scores ranging from 0.94 to 0.98 on the test dataset, which is a translated version of the i2b2 2014 de-identification challenge, across 17 sensitive entities. This level of accuracy is in line with that achieved with manual de-identification by domain experts, suggesting that a fully automated and scalable process is now viable.

We present ArTST, a pre-trained Arabic text and speech transformer for supporting open-source speech technologies for the Arabic language. The model architecture follows the unified-modal framework, SpeechT5, that was recently released for English, and is focused on Modern Standard Arabic (MSA), with plans to extend the model for dialectal and code-switched Arabic in future editions. We pre-trained the model from scratch on MSA speech and text data, and fine-tuned it for the following tasks: Automatic Speech Recognition (ASR), Text-To-Speech synthesis (TTS), and spoken dialect identification. In our experiments comparing ArTST with SpeechT5, as well as with previously reported results in these tasks, ArTST performs on a par with or exceeding the current state-of-the-art in all three tasks. Moreover, we find that our pre-training is conducive for generalization, which is particularly evident in the low-resource TTS task. The pre-trained model as well as the fine-tuned ASR and TTS models are released for research use.

Despite the purported multilingual proficiency of instruction-finetuned large language models (LLMs) such as ChatGPT and Bard, the linguistic inclusivity of these models remains insufficiently explored. Considering this constraint, we present a thorough assessment of Bard and ChatGPT (encompassing both GPT-3.5 and GPT-4) regarding their machine translation proficiencies across ten varieties of Arabic. Our evaluation covers diverse Arabic varieties such as Classical Arabic (CA), Modern Standard Arabic (MSA), and several country-level dialectal variants. Our analysis indicates that LLMs may encounter challenges with dialects for which minimal public datasets exist, but on average are better translators of dialects than existing commercial systems. On CA and MSA, instruction-tuned LLMs, however, trail behind commercial systems such as Google Translate. Finally, we undertake a human-centric study to scrutinize the efficacy of the relatively recent model, Bard, in following human instructions during translation tasks. Our analysis reveals a circumscribed capability of Bard in aligning with human instructions in translation contexts. Collectively, our findings underscore that prevailing LLMs remain far from inclusive, with only limited ability to cater for the linguistic and cultural intricacies of diverse communities.

In this work, we approach the problem of Qur’anic information retrieval (IR) in Arabic and English. Using the latest state-of-the-art methods in neural IR, we research what helps to tackle this task more efficiently. Training retrieval models requires a lot of data, which is difficult to obtain for training in-domain. Therefore, we commence with training on a large amount of general domain data and then continue training on in-domain data. To handle the lack of in-domain data, we employed a data augmentation technique, which considerably improved results in MRR@10 and NDCG@5 metrics, setting the state-of-the-art in Qur’anic IR for both English and Arabic. The absence of an Islamic corpus and domain-specific model for IR task in English motivated us to address this lack of resources and take preliminary steps of the Islamic corpus compilation and domain-specific language model (LM) pre-training, which helped to improve the performance of the retrieval models that use the domain-specific LM as the shared backbone. We examined several language models (LMs) in Arabic to select one that efficiently deals with the Qur’anic IR task. Besides transferring successful experiments from English to Arabic, we conducted additional experiments with retrieval task in Arabic to amortize the scarcity of general domain datasets used to train the retrieval models. Handling Qur’anic IR task combining English and Arabic allowed us to enhance the comparison and share valuable insights across models and languages.

Text summarization has been intensively studied in many languages, and some languages have reached advanced stages. Yet, Arabic Text Summarization (ATS) is still in its developing stages. Existing ATS datasets are either small or lack diversity. We build, LANS, a large-scale and diverse dataset for Arabic Text Summarization task. LANS offers 8.4 million articles and their summaries extracted from newspapers websites’ metadata between 1999 and 2019. The high-quality and diverse summaries are written by journalists from 22 major Arab newspapers and include an eclectic mix of at least more than 7 topics from each source. We conduct an intrinsic evaluation on LANS by both automatic and human evaluations. Human evaluation of 1,000 random samples reports 95.4% accuracy for our collected summaries, and automatic evaluation quantifies the diversity and abstractness of the summaries.

Large language models (LLMs) finetuned to follow human instruction have recently exhibited significant capabilities in various English NLP tasks. However, their performance in grammatical error correction (GEC), especially on languages other than English, remains significantly unexplored. In this work, we evaluate the abilities of instruction finetuned LLMs in Arabic GEC, a complex task due to Arabic’s rich morphology. Our findings suggest that various prompting methods, coupled with (in-context) few-shot learning, demonstrate considerable effectiveness, with GPT-4 achieving up to 65.49 F1 score under expert prompting (approximately 5 points higher than our established baseline). Despite these positive results, we find that instruction finetuned models, regardless of their size, are still outperformed by fully finetuned ones, even if they are significantly smaller in size. This disparity highlights substantial room for improvements for LLMs. Inspired by methods used in low-resource machine translation, we also develop a method exploiting synthetic data that significantly outperforms previous models on two standard Arabic benchmarks. Our best model achieves a new SOTA on Arabic GEC, with 73.29 and 73.26 F1 on the 2014 and 2015 QALB datasets, respectively, compared to peer-reviewed published baselines.

Recent advancements in self-supervised speech-representation learning for automatic speech recognition (ASR) approaches have significantly improved the results on many benchmarks with low-cost data labeling. In this paper, we train two self-supervised frameworks for ASR, namely wav2vec, and data2vec, in which we conduct multiple experiments and analyze their results. Furthermore, we introduce Aswat dataset, which covers multiple genres and features speakers with vocal variety. Aswat contains 732 hours of clean Arabic speech that can be used in the pretraining task for learning latent speech representations, which results in achieving a lower word error rate (WER) in Arabic ASR. We report the baseline results and achieve state-of-the-art WERs of 11.7% and 10.3% on Common Voice (CV) and the second round of Multi-Genre Broadcast (MGB-2) respectively, as a result of including our dataset Aswat.

While significant progress has been made in benchmarking Large Language Models (LLMs) across various tasks, there is a lack of comprehensive evaluation of their abilities in responding to multi-turn instructions in less-commonly tested languages like Arabic. Our paper offers a detailed examination of the proficiency of open LLMs in such scenarios in Arabic. Utilizing a customized Arabic translation of the MT-Bench benchmark suite, we employ GPT-4 as a uniform evaluator for both English and Arabic queries to assess and compare the performance of the LLMs on various open-ended tasks. Our findings reveal variations in model responses on different task categories, e.g., logic vs. literacy, when instructed in English or Arabic. We find that fine-tuned base models using multilingual and multi-turn datasets could be competitive to models trained from scratch on multilingual data. Finally, we hypothesize that an ensemble of small, open LLMs could perform competitively to proprietary LLMs on the benchmark.

In this paper, we study the transferability of Named Entity Recognition (NER) models between Arabic dialects. This question is important because the available manually-annotated resources are not distributed equally across dialects: Modern Standard Arabic (MSA) is much richer than other dialects for which little to no datasets exist. How well does a NER model, trained on MSA, perform on other dialects? To answer this question, we construct four datasets. The first is an MSA dataset extracted from the ACE 2005 corpus. The others are datasets for Egyptian, Morocan and Syrian which we manually annotate following the ACE guidelines. We train a span-based NER model on top of a pretrained language model (PLM) encoder on the MSA data and study its performance on the other datasets in zero-shot settings. We study the performance of multiple PLM encoders from the literature and show that they achieve acceptable performance with no annotation effort. Our annotations and models are publicly available (https://github.com/niamaelkhbir/Arabic-Cross-Dialectal-NER).

Product information in e-commerce is usually localized using machine translation (MT) systems. Arabic language has rich morphology and dialectal variations, so Arabic MT in e-commerce training requires a larger volume of data from diverse data sources; Given the dynamic nature of e-commerce, such data needs to be acquired periodically to update the MT. Consequently, validating the quality of training data periodically within an industrial setting presents a notable challenge. Meanwhile, the performance of MT systems is significantly impacted by the quality and appropriateness of the training data. Hence, this study first examines the Arabic MT in e-commerce and investigates the data quality challenges for English-Arabic MT in e-commerce then proposes heuristics-based and topic-based data selection approaches to improve MT for product information. Both online and offline experiment results have shown our proposed approaches are effective, leading to improved shopping experiences for customers.

Extracting and disambiguating geolocation information from social media data enables effective disaster management, as it helps response authorities; for example, locating incidents for planning rescue activities and affected people for evacuation. Nevertheless, the dearth of resources and tools hinders the development and evaluation of Location Mention Disambiguation (LMD) models in the disaster management domain. Consequently, the LMD task is greatly understudied, especially for the low resource languages such as Arabic. To fill this gap, we introduce IDRISI-D, the largest to date English and the first Arabic public LMD datasets. Additionally, we introduce a modified hierarchical evaluation framework that offers a lenient and nuanced evaluation of LMD systems. We further benchmark IDRISI-D datasets using representative baselines and show the competitiveness of BERT-based models.

We present CamelParser2.0, an open-source Python-based Arabic dependency parser targeting two popular Arabic dependency formalisms, the Columbia Arabic Treebank (CATiB), and Universal Dependencies (UD). The CamelParser2.0 pipeline handles the processing of raw text and produces tokenization, part-of-speech and rich morphological features. As part of developing CamelParser2.0, we explore many system design hyper-parameters, such as parsing model architecture and pretrained language model selection, achieving new state-of-the-art performance across diverse Arabic genres under gold and predicted tokenization settings.

Bilingual Lexical Induction (BLI) is a core challenge in NLP, it relies on the relative isomorphism of individual embedding spaces. Existing attempts aimed at controlling the relative isomorphism of different embedding spaces fail to incorporate the impact of semantically related words in the model training objective. To address this, we propose GARI that combines the distributional training objectives with multiple isomorphism losses guided by the graph attention network. GARI considers the impact of semantical variations of words in order to define the relative isomorphism of the embedding spaces. Experimental evaluation using the Arabic language data set shows that GARI outperforms the existing research by improving the average P@1 by a relative score of up to 40.95% and 76.80% for in-domain and domain mismatch settings respectively.

We present ArTrivia, a new Arabic question-answering dataset consisting of more than 10,000 question-answer pairs along with relevant passages, covering a wide range of 18 diverse topics in Arabic. We created our dataset using a newly proposed pipeline that leverages diverse structured data sources from Arabic Wikipedia. Moreover, we conducted a comprehensive statistical analysis of ArTrivia and assessed the performance of each component in our pipeline. Additionally, we compared the performance of ArTrivia against the existing TyDi QA dataset using various experimental setups. Our analysis highlights the significance of often overlooked aspects in dataset creation, such as answer normalization, in enhancing the quality of QA datasets. Our evaluation also shows that ArTrivia presents more challenging and out-of-distribution questions to TyDi, raising questions about the feasibility of using ArTrivia as a complementary dataset to TyDi.

In digital communication, emoji are essential in decoding nuances such as irony, sarcasm, and humour. However, their incorporation in Arabic natural language processing (NLP) has been cautious because of the perceived complexities of the Arabic language. This paper introduces ArSarcasMoji, a dataset of 24,630 emoji-augmented texts, with 17. 5% that shows irony. Through our analysis, we highlight specific emoji patterns paired with sentiment roles that denote irony in Arabic texts. The research counters prevailing notions, emphasising the importance of emoji’s role in understanding Arabic textual irony, and addresses their potential for accurate irony detection in Arabic digital content.

Wikipedia articles are a widely used source of training data for Natural Language Processing (NLP) research, particularly as corpora for low-resource languages like Arabic. However, it is essential to understand the extent to which these corpora reflect the representative contributions of native speakers, especially when many entries in a given language are directly translated from other languages or automatically generated through automated mechanisms. In this paper, we study the performance implications of using inorganic corpora that are not representative of native speakers and are generated through automated techniques such as bot generation or automated template-based translation. The case of the Arabic Wikipedia editions gives a unique case study of this since the Moroccan Arabic Wikipedia edition (ARY) is small but representative, the Egyptian Arabic Wikipedia edition (ARZ) is large but unrepresentative, and the Modern Standard Arabic Wikipedia edition (AR) is both large and more representative. We intrinsically evaluate the performance of two main NLP upstream tasks, namely word representation and language modeling, using word analogy evaluations and fill-mask evaluations using our two newly created datasets: Arab States Analogy Dataset (ASAD) and Masked Arab States Dataset (MASD). We demonstrate that for good NLP performance, we need both large and organic corpora; neither alone is sufficient. We show that producing large corpora through automated means can be a counter-productive, producing models that both perform worse and lack cultural richness and meaningful representation of the Arabic language and its native speakers.

Understanding Arabic text and generating human-like responses is a challenging task. While many researchers have proposed models and solutions for individual problems, there is an acute shortage of a comprehensive Arabic natural language generation toolkit that is capable of handling a wide range of tasks. In this work, we present a robust Arabic text-to-text Transformer model, namely AraT5v2, methodically trained on extensive and diverse data, utilizing an extended sequence length of 2,048 tokens. We explore various pretraining strategies including unsupervised, supervised, and joint pertaining, under both single and multitask settings. Our models outperform competitive baselines with large margins. We take our work one step further by developing and publicly releasing OCTOPUS, a Python-based package and command-line toolkit tailored for eight Arabic generation tasks all exploiting a single model. We provide a link to the models and the toolkit through our public repository.

Recent advances in the space of Arabic large language models have opened up a wealth of potential practical applications. From optimal training strategies, large scale data acquisition and continuously increasing NLP resources, the Arabic LLM landscape has improved in a very short span of time, despite being plagued by training data scarcity and limited evaluation resources compared to English. In line with contributing towards this ever-growing field, we introduce AlGhafa, a new multiple-choice evaluation benchmark for Arabic LLMs. For showcasing purposes, we train a new suite of models, including a 14 billion parameter model, the largest monolingual Arabic decoder-only model to date. We use a collection of publicly available datasets, as well as a newly introduced HandMade dataset consisting of 8 billion tokens. Finally, we explore the quantitative and qualitative toxicity of several Arabic models, comparing our models to existing public Arabic LLMs.

This paper presents the ArBanking77, a large Arabic dataset for intent detection in the banking domain. Our dataset was arabized and localized from the original English Banking77 dataset, which consists of 13,083 queries to ArBanking77 dataset with 31,404 queries in both Modern Standard Arabic (MSA) and Palestinian dialect, with each query classified into one of the 77 classes (intents). Furthermore, we present a neural model, based on AraBERT, fine-tuned on ArBanking77, which achieved an F1-score of 0.9209 and 0.8995 on MSA and Palestinian dialect, respectively. We performed extensive experimentation in which we simulated low-resource settings, where the model is trained on a subset of the data and augmented with noisy queries to simulate colloquial terms, mistakes and misspellings found in real NLP systems, especially live chat queries. The data and the models are publicly available at https://sina.birzeit.edu/arbanking77.

This paper presents the first Arabic crossword puzzle generator driven by advanced AI technology. Leveraging cutting-edge large language models including GPT4, GPT3-Davinci, GPT3-Curie, GPT3-Babbage, GPT3-Ada, and BERT, the system generates distinctive and challenging clues. Based on a dataset comprising over 50,000 clue-answer pairs, the generator employs fine-tuning, few/zero-shot learning strategies, and rigorous quality-checking protocols to enforce the generation of high-quality clue-answer pairs. Importantly, educational crosswords contribute to enhancing memory, expanding vocabulary, and promoting problem-solving skills, thereby augmenting the learning experience through a fun and engaging approach, reshaping the landscape of traditional learning methods. The overall system can be exploited as a powerful educational tool that amalgamates AI and innovative learning techniques, heralding a transformative era for Arabic crossword puzzles and the intersection of technology and education.

Research studies on Machine Translation (MT) between Modern Standard Arabic (MSA) and English are abundant. However, studies on MT between Omani Arabic (OA) dialects and English are very scarce. This research study focuses on the lack of availability of an Omani dialect parallel dataset, as well as MT of OA to English. The study uses social media data from X (formerly Twitter) to build an authentic parallel text of the Omani dialects. The research presents baseline results on this dataset using Google Translate, Microsoft Translation, and Marian NMT. A taxonomy of the most common linguistic errors is used to analyze the translations made by the NMT systems to provide insights on future improvements. Finally, transfer learning is used to adapt Marian NMT to the Omani dialect, which significantly improved by 9.88 points in the BLEU score.

Traditional NER systems are typically trained to recognize coarse-grained categories of entities, and less attention is given to classifying entities into a hierarchy of fine-grained lower-level sub-types. This article aims to advance Arabic NER with fine-grained entities. We chose to extend Wojood (an open-source Nested Arabic Named Entity Corpus) with sub-types. In particular, four main entity types in Wojood (geopolitical entity (GPE), location (LOC), organization (ORG), and facility (FAC) are extended with 31 sub-types of entities. To do this, we first revised Wojood’s annotations of GPE, LOC, ORG, and FAC to be compatible with the LDC’s ACE guidelines, which yielded 5, 614 changes. Second, all mentions of GPE, LOC, ORG, and FAC (~ 44K) in Wojood are manually annotated with the LDC’s ACE subtypes. This extended version of Wojood is called WojoodFine. To evaluate our annotations, we measured the inter-annotator agreement (IAA) using both Cohen’s Kappa and F1 score, resulting in 0.9861 and 0.9889, respectively. To compute the baselines of WojoodFine, we fine-tune three pre-trained Arabic BERT encoders in three settings: flat NER, nested NER and nested NER with sub-types and achieved F1 score of 0.920, 0.866, and 0.885, respectively. Our corpus and models are open source and available at https://sina.birzeit.edu/wojood/.

Numerous languages exhibit shared characteristics, especially in morphological features. For instance, Arabic and Russian both belong to the fusional language category. The question arises: Do such common traits influence language comprehension across diverse linguistic backgrounds? This study explores the possibility of transferring comprehension skills across languages to Arabic in a zero-shot scenario. Specifically, we demonstrate that training language models on other languages can enhance comprehension of Arabic, as evidenced by our evaluations in three key tasks: natural language inference, question answering, and named entity recognition. Our experiments reveal that certain morphologically rich languages (MRLs), such as Russian, display similarities to Arabic when assessed in a zero-shot context, particularly in tasks like question answering and natural language inference. However, this similarity is less pronounced in tasks like named entity recognition.

Large Language Models (LLMs) such as ChatGPT and Bard AI have gained much attention due to their outstanding performance on a range of NLP tasks. These models have demonstrated remarkable proficiency across various languages without the necessity for full supervision. Nevertheless, their performance in low-resource languages and dialects, like Arabic dialects in comparison to English, remains to be investigated. In this paper, we conduct a comprehensive evaluation of three LLMs for Dialectal Arabic Sentiment Analysis: namely, ChatGPT based on GPT-3.5 and GPT-4, and Bard AI. We use a Saudi dialect Twitter dataset to assess their capability in sentiment text classification and generation. For classification, we compare the performance of fully fine-tuned Arabic BERT-based models with the LLMs in few-shot settings. For data generation, we evaluate the quality of the generated new sentiment samples using human and automatic evaluation methods. The experiments reveal that GPT-4 outperforms GPT-3.5 and Bard AI in sentiment analysis classification, rivaling the top-performing fully supervised BERT-based language model. However, in terms of data generation, compared to manually annotated authentic data, these generative models often fall short in producing high-quality Dialectal Arabic text suitable for sentiment analysis.

Delegating short answer grading to automated systems enhances efficiency, giving teachers more time for vital human-centered aspects of education. Studies in automatic short answer grading (ASAG) approach the problem from instance-based or reference-based perspectives. Recent studies have favored instance-based methods, but they demand substantial data for training, which is often scarce in classroom settings. This study compares both approaches using an Arabic ASAG dataset. We employ in-context meta-learning for instance-based and semantic score-based similarity for reference-based grading. Results show both methods outperform a baseline and occasionally even surpass human raters when grading unseen answers. Notably, the semantic score-based similarity approach excels in zero-shot settings, outperforming in-context meta-learning. Our work contributes insights to Arabic ASAG and introduces a prompt category classification model, leveraging GPT3.5 to augment Arabic data for improved performance.

SALMA, the first Arabic sense-annotated corpus, consists of ~34K tokens, which are all sense-annotated. The corpus is annotated using two different sense inventories simultaneously (Modern and Ghani). SALMA novelty lies in how tokens and senses are associated. Instead of linking a token to only one intended sense, SALMA links a token to multiple senses and provides a score to each sense. A smart web-based annotation tool was developed to support scoring multiple senses against a given word. In addition to sense annotations, we also annotated the corpus using six types of named entities. The quality of our annotations was assessed using various metrics (Kappa, Linear Weighted Kappa, Quadratic Weighted Kappa, Mean Average Error, and Root Mean Square Error), which show very high inter-annotator agreement. To establish a Word Sense Disambiguation baseline using our SALMA corpus, we developed an end-to-end Word Sense Disambiguation system using Target Sense Verification. We used this system to evaluate three Target Sense Verification models available in the literature. Our best model achieved an accuracy with 84.2% using Modern and 78.7% using Ghani. The full corpus and the annotation tool are open-source and publicly available at https://sina.birzeit.edu/salma/.

This paper provides a systematic analysis and comparison of the performance of state-of-the-art models on the task of fine-grained Arabic dialect identification using the MADAR parallel corpus. We test approaches based on pre-trained transformer language models in addition to Naive Bayes models with a rich set of various features. Through a comprehensive data- and error analysis, we provide valuable insights into the strengths and weaknesses of both approaches. We discuss which dialects are more challenging to differentiate, and identify potential sources of errors. Our analysis reveals an important problem with identical sentences across dialect classes in the test set of the MADAR-26 corpus, which may confuse any classifier. We also show that none of the tested approaches captures the subtle distinctions between closely related dialects.

Automatic Arabic Dialect Identification (ADI) of text has gained great popularity since it was introduced in the early 2010s. Multiple datasets were developed, and yearly shared tasks have been running since 2018. However, ADI systems are reported to fail in distinguishing between the micro-dialects of Arabic. We argue that the currently adopted framing of the ADI task as a single-label classification problem is one of the main reasons for that. We highlight the limitation of the incompleteness of the Dialect labels and demonstrate how it impacts the evaluation of ADI systems. A manual error analysis for the predictions of an ADI, performed by 7 native speakers of different Arabic dialects, revealed that ≈ 67% of the validated errors are not true errors. Consequently, we propose framing ADI as a multi-label classification task and give recommendations for designing new ADI datasets.

Most recent models for Arabic topic classification leveraged fine-tuning existing pre-trained transformer models and targeted a limited number of categories. More recently, advances in automated ML and generative models introduced novel potentials for the task. While these approaches work for English, it is a question of whether they perform well for low-resourced languages; Arabic in particular. This paper presents (i) ArBoNeClass; a novel Arabic dataset with an extended 14-topic class set covering modern books from social sciences and humanities along with newspaper articles, and (ii) a set of topic classifiers built from it. We finetuned an open LLM model to build ArGTClass. We compared its performance against the best models built with Vertex AI (Google), AutoML(H2O), and AutoTrain(HuggingFace). ArGTClass outperformed the VertexAi and AutoML models and was reasonably similar to the AutoTrain model.