We consider an unanswered question in the discourse processing community: why do relation classifiers trained on explicit examples (with connectives removed) perform poorly in real implicit scenarios? Prior work claimed this is due to linguistic dissimilarity between explicit and implicit examples but provided no empirical evidence. In this study, we show that one cause for such failure is a label shift after connectives are eliminated. Specifically, we find that the discourse relations expressed by some explicit instances will change when connectives disappear. Unlike previous work manually analyzing a few examples, we present empirical evidence at the corpus level to prove the existence of such shift. Then, we analyze why label shift occurs by considering factors such as the syntactic role played by connectives, ambiguity of connectives, and more. Finally, we investigate two strategies to mitigate the label shift: filtering out noisy data and joint learning with connectives. Experiments on PDTB 2.0, PDTB 3.0, and the GUM dataset demonstrate that classifiers trained with our strategies outperform strong baselines.
Finding evidence for claims from content presented in experimental results of scientific articles is difficult. The evidence is often presented in the form of tables and figures, and correctly matching it to scientific claims presents automation challenges. The Context24 shared task is launched to support the development of systems able to verify claims by extracting supporting evidence from articles. We explore different facets of this shared task modelled as a search problem and as an information extraction task. We experiment with a range of methods in each of these categories for the two sub-tasks of evidence identification and grounding context identification in the Context24 shared task.
Prompt-based usage of Large Language Models (LLMs) is an increasingly popular way to tackle many well-known natural language problems. This trend is due, in part, to the appeal of the In-Context Learning (ICL) prompt set-up, in which a few selected training examples are provided along with the inference request. ICL, a type of few-shot learning, is especially attractive for natural language processing (NLP) tasks defined for specialised domains, such as entity extraction from scientific documents, where the annotation is very costly due to expertise requirements for the annotators. In this paper, we present a comprehensive analysis of in-context sample selection methods for entity extraction from scientific documents using GPT-3.5 and compare these results against a fully supervised transformer-based baseline. Our results indicate that the effectiveness of the in-context sample selection methods is heavily domain-dependent, but the improvements are more notable for problems with a larger number of entity types. More in-depth analysis shows that ICL is more effective for low-resource set-ups of scientific information extraction
Long document summarisation, a challenging summarisation scenario, is the focus of the recently proposed LongSumm shared task. One of the limitations of this shared task has been its use of a single family of metrics for evaluation (the ROUGE metrics). In contrast, other fields, like text generation, employ multiple metrics. We replicated the LongSumm evaluation using multiple test set samples (vs. the single test set of the official shared task) and investigated how different metrics might complement each other in this evaluation framework. We show that under this more rigorous evaluation, (1) some of the key learnings from Longsumm 2020 and 2021 still hold, but the relative ranking of systems changes, and (2) the use of additional metrics reveals additional high-quality summaries missed by ROUGE, and (3) we show that SPICE is a candidate metric for summarisation evaluation for LongSumm.
This paper focuses on Seq2Seq (S2S) constrained text generation where the text generator is constrained to mention specific words which are inputs to the encoder in the generated outputs. Pre-trained S2S models or a Copy Mechanism are trained to copy the surface tokens from encoders to decoders, but they cannot guarantee constraint satisfaction. Constrained decoding algorithms always produce hypotheses satisfying all constraints. However, they are computationally expensive and can lower the generated text quality. In this paper, we propose Mention Flags (MF), which traces whether lexical constraints are satisfied in the generated outputs in an S2S decoder. The MF models can be trained to generate tokens in a hypothesis until all constraints are satisfied, guaranteeing high constraint satisfaction. Our experiments on the Common Sense Generation task (CommonGen) (Lin et al., 2020), End2end Restaurant Dialog task (E2ENLG) (Duˇsek et al., 2020) and Novel Object Captioning task (nocaps) (Agrawal et al., 2019) show that the MF models maintain higher constraint satisfaction and text quality than the baseline models and other constrained decoding algorithms, achieving state-of-the-art performance on all three tasks. These results are achieved with a much lower run-time than constrained decoding algorithms. We also show that the MF models work well in the low-resource setting.
Relation prediction informed from a combination of text corpora and curated knowledge bases, combining knowledge graph completion with relation extraction, is a relatively little studied task. A system that can perform this task has the ability to extend an arbitrary set of relational database tables with information extracted from a document corpus. OpenKi[1] addresses this task through extraction of named entities and predicates via OpenIE tools then learning relation embeddings from the resulting entity-relation graph for relation prediction, outperforming previous approaches. We present an extension of OpenKi that incorporates embeddings of text-based representations of the entities and the relations. We demonstrate that this results in a substantial performance increase over a system without this information.
Novel Object Captioning is a zero-shot Image Captioning task requiring describing objects not seen in the training captions, but for which information is available from external object detectors. The key challenge is to select and describe all salient detected novel objects in the input images. In this paper, we focus on this challenge and propose the ECOL-R model (Encouraging Copying of Object Labels with Reinforced Learning), a copy-augmented transformer model that is encouraged to accurately describe the novel object labels. This is achieved via a specialised reward function in the SCST reinforcement learning framework (Rennie et al., 2017) that encourages novel object mentions while maintaining the caption quality. We further restrict the SCST training to the images where detected objects are mentioned in reference captions to train the ECOL-R model. We additionally improve our copy mechanism via Abstract Labels, which transfer knowledge from known to novel object types, and a Morphological Selector, which determines the appropriate inflected forms of novel object labels. The resulting model sets new state-of-the-art on the nocaps (Agrawal et al., 2019) and held-out COCO (Hendricks et al., 2016) benchmarks.
Vent is a specialised iOS/Android social media platform with the stated goal to encourage people to post about their feelings and explicitly label them. In this paper, we study a snapshot of more than 100 million messages obtained from the developers of Vent, together with the labels assigned by the authors of the messages. We establish the quality of the self-annotated data by conducting a qualitative analysis, a vocabulary based analysis, and by training and testing an emotion classifier. We conclude that the self-annotated labels of our corpus are indeed indicative of the emotional contents expressed in the text and thus can support more detailed analyses of emotion expression on social media, such as emotion trajectories and factors influencing them.
For many NLP applications of online reviews, comparison of two opinion-bearing sentences is key. We argue that, while general purpose text similarity metrics have been applied for this purpose, there has been limited exploration of their applicability to opinion texts. We address this gap in the literature, studying: (1) how humans judge the similarity of pairs of opinion-bearing sentences; and, (2) the degree to which existing text similarity metrics, particularly embedding-based ones, correspond to human judgments. We crowdsourced annotations for opinion sentence pairs and our main findings are: (1) annotators tend to agree on whether or not opinion sentences are similar or different; and (2) embedding-based metrics capture human judgments of “opinion similarity” but not “opinion difference”. Based on our analysis, we identify areas where the current metrics should be improved. We further propose to learn a similarity metric for opinion similarity via fine-tuning the Sentence-BERT sentence-embedding network based on review text and weak supervision by review ratings. Experiments show that our learned metric outperforms existing text similarity metrics and especially show significantly higher correlations with human annotations for differing opinions.
There are many different ways in which external information might be used in a NLP task. This paper investigates how external syntactic information can be used most effectively in the Semantic Role Labeling (SRL) task. We evaluate three different ways of encoding syntactic parses and three different ways of injecting them into a state-of-the-art neural ELMo-based SRL sequence labelling model. We show that using a constituency representation as input features improves performance the most, achieving a new state-of-the-art for non-ensemble SRL models on the in-domain CoNLL’05 and CoNLL’12 benchmarks.
One of the most common metrics to automatically evaluate opinion summaries is ROUGE, a metric developed for text summarisation. ROUGE counts the overlap of word or word units between a candidate summary against reference summaries. This formulation treats all words in the reference summary equally. In opinion summaries, however, not all words in the reference are equally important. Opinion summarisation requires to correctly pair two types of semantic information: (1) aspect or opinion target; and (2) polarity of candidate and reference summaries. We investigate the suitability of ROUGE for evaluating opin-ion summaries of online reviews. Using three simulation-based experiments, we evaluate the behaviour of ROUGE for opinion summarisation on the ability to match aspect and polarity. We show that ROUGE cannot distinguish opinion summaries of similar or opposite polarities for the same aspect. Moreover,ROUGE scores have significant variance under different configuration settings. As a result, we present three recommendations for future work that uses ROUGE to evaluate opinion summarisation.
In social media, demographic inference is a critical task in order to gain a better understanding of a cohort and to facilitate interacting with one’s audience. Most previous work has made independence assumptions over topological, textual and label information on social networks. In this work, we employ recursive neural networks to break down these independence assumptions to obtain inference about demographic characteristics on Twitter. We show that our model performs better than existing models including the state-of-the-art.
In this paper, we describe CSIRO Data61’s participation in the Geolocation shared task at the Workshop for Noisy User-generated Text. Our approach was to use ensemble methods to capitalise on four component methods: heuristics based on metadata, a label propagation method, timezone text classifiers, and an information retrieval approach. The ensembles we explored focused on examining the role of language technologies in geolocation prediction and also in examining the use of hard voting and cascading ensemble methods. Based on the accuracy of city-level predictions, our systems were the best performing submissions at this year’s shared task. Furthermore, when estimating the latitude and longitude of a user, our median error distance was accurate to within 30 kilometers.