The NLI4CT task assesses Natural Language Inference systems in predicting whether hypotheses entail or contradict evidence from Clinical Trial Reports. In this study, we evaluate various Large Language Models (LLMs) with multiple strategies, including Chain-of-Thought, In-Context Learning, and Parameter-Efficient Fine-Tuning (PEFT). We propose a PEFT method to improve the consistency of LLMs by merging adapters that were fine-tuned separately using triplet and language modelling objectives. We found that merging the two PEFT adapters improves the F1 score (+0.0346) and consistency (+0.152) of the LLMs. However, our novel methods did not produce more accurate results than GPT-4 in terms of faithfulness and consistency. Averaging the three metrics, GPT-4 ranks joint-first in the competition with 0.8328. Finally, our contamination analysis with GPT-4 indicates that there was no test data leakage. Our code is available at https://github.com/EdinburghClinicalNLP/semeval_nli4ct.
Adapting pretrained language models to novel domains, such as clinical applications, traditionally involves retraining their entire set of parameters. Parameter-Efficient Fine-Tuning (PEFT) techniques for fine-tuning language models significantly reduce computational requirements by selectively fine-tuning small subsets of parameters. In this study, we propose a two-step PEFT framework and evaluate it in the clinical domain. Our approach combines a specialised PEFT adapter layer designed for clinical domain adaptation with another adapter specialised for downstream tasks. We evaluate the framework on multiple clinical outcome prediction datasets, comparing it to clinically trained language models. Our framework achieves a better AUROC score averaged across all clinical downstream tasks compared to clinical language models. In particular, we observe large improvements of 4-5% AUROC in large-scale multilabel classification tasks, such as diagnoses and procedures classification. To our knowledge, this study is the first to provide an extensive empirical analysis of the interplay between PEFT techniques and domain adaptation in an important real-world domain of clinical applications.
The MEDIQA-CORR 2024 shared task aims to assess the ability of Large Language Models (LLMs) to identify and correct medical errors in clinical notes. In this study, we evaluate the capability of general LLMs, specifically GPT-3.5 and GPT-4, to identify and correct medical errors with multiple prompting strategies. Recognising the limitation of LLMs in generating accurate corrections only via prompting strategies, we propose incorporating error-span predictions from a smaller, fine-tuned model in two ways: 1) by presenting it as a hint in the prompt and 2) by framing it as multiple-choice questions from which the LLM can choose the best correction. We found that our proposed prompting strategies significantly improve the LLM’s ability to generate corrections. Our best-performing solution with 8-shot + CoT + hints ranked sixth in the shared task leaderboard. Additionally, our comprehensive analyses show the impact of the location of the error sentence, the prompted role, and the position of the multiple-choice option on the accuracy of the LLM. This prompts further questions about the readiness of LLM to be implemented in real-world clinical settings.
This paper presents an overview of text visualization techniques relevant for data perspectivism, aiming to facilitate analysis of annotated datasets for the datasets’ creators and stakeholders. Data perspectivism advocates for publishing non-aggregated, annotated text data, recognizing that for highly subjective tasks, such as bias detection and hate speech detection, disagreements among annotators may indicate conflicting yet equally valid interpretations of a text. While the publication of non-aggregated, annotated data makes different interpretations of text corpora available, barriers still exist to investigating patterns and outliers in annotations of the text. Techniques from text visualization can overcome these barriers, facilitating intuitive data analysis for NLP researchers and practitioners, as well as stakeholders in NLP systems, who may not have data science or computing skills. In this paper we discuss challenges with current dataset creation practices and annotation platforms, followed by a discussion of text visualization techniques that enable open-ended, multi-faceted, and iterative analysis of annotated data.
Mitigating harms from gender biased language in Natural Language Processing (NLP) systems remains a challenge, and the situated nature of language means bias is inescapable in NLP data. Though efforts to mitigate gender bias in NLP are numerous, they often vaguely define gender and bias, only consider two genders, and do not incorporate uncertainty into models. To address these limitations, in this paper we present a taxonomy of gender biased language and apply it to create annotated datasets. We created the taxonomy and annotated data with the aim of making gender bias in language transparent. If biases are communicated clearly, varieties of biased language can be better identified and measured. Our taxonomy contains eleven types of gender biases inclusive of people whose gender expressions do not fit into the binary conceptions of woman and man, and whose gender differs from that they were assigned at birth, while also allowing annotators to document unknown gender information. The taxonomy and annotated data will, in future work, underpin analysis and more equitable language model development.
Like most other minority languages, Scottish Gaelic has limited tools and resources available for Natural Language Processing research and applications. These limitations restrict the potential of the language to participate in modern speech technology, while also restricting research in fields such as corpus linguistics and the Digital Humanities. At the same time, Gaelic has a long written history, is well-described linguistically, and is unusually well-supported in terms of potential NLP training data. For instance, archives such as the School of Scottish Studies hold thousands of digitised recordings of vernacular speech, many of which have been transcribed as paper-based, handwritten manuscripts. In this paper, we describe a project to digitise and recognise a corpus of handwritten narrative transcriptions, with the intention of re-purposing it to develop a Gaelic speech recognition system.
This paper discusses our efforts to develop a full automatic speech recognition (ASR) system for Scottish Gaelic, starting from a point of limited resource. Building ASR technology is important for documenting and revitalising endangered languages; it enables existing resources to be enhanced with automatic subtitles and transcriptions, improves accessibility for users, and, in turn, encourages continued use of the language. In this paper, we explain the many difficulties faced when collecting minority language data for speech recognition. A novel cross-lingual approach to the alignment of training data is used to overcome one such difficulty, and in this way we demonstrate how majority language resources can bootstrap the development of lower-resourced language technology. We use the Kaldi speech recognition toolkit to develop several Gaelic ASR systems, and report a final WER of 26.30%. This is a 9.50% improvement on our original model.
Medical document coding is the process of assigning labels from a structured label space (ontology – e.g., ICD-9) to medical documents. This process is laborious, costly, and error-prone. In recent years, efforts have been made to automate this process with neural models. The label spaces are large (in the order of thousands of labels) and follow a big-head long-tail label distribution, giving rise to few-shot and zero-shot scenarios. Previous efforts tried to address these scenarios within the model, leading to improvements on rare labels, but worse results on frequent ones. We propose data augmentation and synthesis techniques in order to address these scenarios. We further introduce an analysis technique for this setting inspired by confusion matrices. This analysis technique points to the positive impact of data augmentation and synthesis, but also highlights more general issues of confusion within families of codes, and underprediction.
This paper reports on the performance of Edinburgh_UCL_Health’s models in the Social Media Mining for Health (SMM4H) 2022 shared tasks. Our team participated in the tasks related to the Identification of Adverse Drug Events (ADEs), the classification of change in medication (change-med) and the classification of self-report of vaccination (self-vaccine). Our best performing models are based on DeepADEMiner (with respective F1= 0.64, 0.62 and 0.39 for ADE identification), on a GloVe model trained on Twitter (with F1=0.11 for the change-med) and finally on a stack embedding including a layer of Glove embedding and two layers of Flair embedding (with F1= 0.77 for self-report).
In this paper we provide an account of how we ported a text and data mining course online in summer 2020 as a result of the COVID-19 pandemic and how we improved it in a second pilot run. We describe the course, how we adapted it over the two pilot runs and what teaching techniques we used to improve students’ learning and community building online. We also provide information on the relentless feedback collected during the course which helped us to adapt our teaching from one session to the next and one pilot to the next. We discuss the lessons learned and promote the use of innovative teaching techniques applied to the digital such as digital badges and pair programming in break-out rooms for teaching Natural Language Processing courses to beginners and students with different backgrounds.
Large-Scale Multi-Label Text Classification (LMTC) includes tasks with hierarchical label spaces, such as automatic assignment of ICD-9 codes to discharge summaries. Performance of models in prior art is evaluated with standard precision, recall, and F1 measures without regard for the rich hierarchical structure. In this work we argue for hierarchical evaluation of the predictions of neural LMTC models. With the example of the ICD-9 ontology we describe a structural issue in the representation of the structured label space in prior art, and propose an alternative representation based on the depth of the ontology. We propose a set of metrics for hierarchical evaluation using the depth-based representation. We compare the evaluation scores from the proposed metrics with previously used metrics on prior art LMTC models for ICD-9 coding in MIMIC-III. We also propose further avenues of research involving the proposed ontological representation.
In this paper we describe our attempt to increase the amount of information that can be retrieved through active learning sessions compared to previous approaches. We optimise the annotator’s labelling process using active learning in the context of coreference resolution. Using simulated active learning experiments, we suggest three adjustments to ensure the labelling time is spent as efficiently as possible. All three adjustments provide more information to the machine learner than the baseline, though a large impact on the F1 score over time is not observed. Compared to previous models, we report a marginal F1 improvement on the final coreference models trained using for two out of the three approaches tested when applied to the English OntoNotes 2012 Coreference Resolution data. Our best-performing model achieves 58.01 F1, an increase of 0.93 F1 over the baseline model.
This paper describes work in progress on devising automatic and parallel methods for geoparsing large digital historical textual data by combining the strengths of three natural language processing (NLP) tools, the Edinburgh Geoparser, spaCy and defoe, and employing different tokenisation and named entity recognition (NER) techniques. We apply these tools to a large collection of nineteenth century Scottish geographical dictionaries, and describe preliminary results obtained when processing this data.
We propose a bias-aware methodology to engage with power relations in natural language processing (NLP) research. NLP research rarely engages with bias in social contexts, limiting its ability to mitigate bias. While researchers have recommended actions, technical methods, and documentation practices, no methodology exists to integrate critical reflections on bias with technical NLP methods. In this paper, after an extensive and interdisciplinary literature review, we contribute a bias-aware methodology for NLP research. We also contribute a definition of biased text, a discussion of the implications of biased NLP systems, and a case study demonstrating how we are executing the bias-aware methodology in research on archival metadata descriptions.
We present an in-depth comparison of three clinical information extraction (IE) systems designed to perform entity recognition and negation detection on brain imaging reports: EdIE-R, a bespoke rule-based system, and two neural network models, EdIE-BiLSTM and EdIE-BERT, both multi-task learning models with a BiLSTM and BERT encoder respectively. We compare our models both on an in-sample and an out-of-sample dataset containing mentions of stroke findings and draw on our error analysis to suggest improvements for effective annotation when building clinical NLP models for a new domain. Our analysis finds that our rule-based system outperforms the neural models on both datasets and seems to generalise to the out-of-sample dataset. On the other hand, the neural models do not generalise negation to the out-of-sample dataset, despite metrics on the in-sample dataset suggesting otherwise.
Twitter-related studies often need to geo-locate Tweets or Twitter users, identifying their real-world geographic locations. As tweet-level geotagging remains rare, most prior work exploited tweet content, timezone and network information to inform geolocation, or else relied on off-the-shelf tools to geolocate users from location information in their user profiles. However, such user location metadata is not consistently structured, causing such tools to fail regularly, especially if a string contains multiple locations, or if locations are very fine-grained. We argue that user profile location (UPL) and tweet location need to be treated as distinct types of information from which differing inferences can be drawn. Here, we apply geoparsing to UPLs, and demonstrate how task performance can be improved by adapting our Edinburgh Geoparser, which was originally developed for processing English text. We present a detailed evaluation method and results, including inter-coder agreement. We demonstrate that the optimised geoparser can effectively extract and geo-reference multiple locations at different levels of granularity with an F1-score of around 0.90. We also illustrate how geoparsed UPLs can be exploited for international information trade studies and country-level sentiment analysis.
This paper discusses the problem of utilising multiply annotated data in training biomedical information extraction systems. Two corpora, annotated with entities and relations, and containing a number of multiply annotated documents, are used to train named entity recognition and relation extraction systems. Several methods of automatically combining the multiple annotations to produce a single annotation are compared, but none produces better results than simply picking one of the annotated versions at random. It is also shown that adding extra singly annotated documents produces faster performance gains than adding extra multiply annotated documents.
The influence of English as a global language continues to grow to an extent that its words and expressions permeate the original forms of other languages. This paper evaluates a modular Web-based sub-component of an existing English inclusion classifier and compares it to a corpus-based lookup technique. Both approaches are evaluated on a German gold standard data set. It is demonstrated to what extent the Web-based approach benefits from the amount of data available online and the fact that this data is constantly updated.
In this paper we discuss five different corpora annotated forprotein names. We present several within- and cross-dataset proteintagging experiments showing that different annotation schemes severelyaffect the portability of statistical protein taggers. By means of adetailed error analysis we identify crucial annotation issues thatfuture annotation projects should take into careful consideration.