Reliable multilingual evaluation is difficult, and culturally appropriate evaluation is even harder to achieve.A common practice to fill this gap is to machine-translate English evaluation sets. However, translation introduces language bias and carries over cultural and regional assumptions from the original questions – often testing knowledge irrelevant to the target audience. In this work, we highlight the extent and impact of these biases and present a multilingual evaluation framework that aims to mitigate them through improved translations and annotation practices.Through a large-scale study involving professional and community translators and annotators, we show that state-of-the-art models excel primarily by learning Western-centric concepts. Notably, we find that model rankings on the full MMLU change when evaluated on a subset of questions explicitly marked as culturally sensitive.We release Global MMLU, a multilingual extension of MMLU across 42 languages, featuring improved translation quality, expanded language coverage, and designated subsets labeled as culturally sensitive and culturally agnostic to enable a more comprehensive and equitable benchmark for evaluating language models across diverse linguistic and cultural contexts.

As the use of large language models becomes ever more prevalent, understanding their persuasive abilities, both in ways that can be beneficial and harmful to humans, proves an important task. Previous work has focused on persuasion in the context of negotiations, political debate and advertising. We instead shift the focus to a more realistic setup of a dialogue between a persuadee with an everyday dilemma (e.g., whether to switch to a vegan diet or not) and a persuader with no prior knowledge about the persuadee who is trying to persuade them towards a certain decision based on arguments they feel would be most suited to the persuadee’s persona. We collect and analyze conversations between a human persuadee and either a human persuader or an LLM persuader based on GPT-4. We find that, in this setting, GPT-4 is perceived as both more persuasive and more empathetic, whereas humans are more skilled at discovering new information about the person they are speaking to. This research provides the groundwork for future work predicting the persuasiveness of utterances in conversation across a range of topics.

Identification of stance has recently gained a lot of attention with the extreme growth of fake news and filter bubbles. Over the last decade, many feature-based and deep-learning approaches have been proposed to solve stance detection. However, almost none of the existing works focus on providing a meaningful explanation for their prediction. In this work, we study stance detection with an emphasis on generating explanations for the predicted stance by capturing the pivotal argumentative structure embedded in a document. We propose to build a stance tree that utilizes rhetorical parsing to construct an evidence tree and to use Dempster Shafer Theory to aggregate the evidence. Human studies show that our unsupervised technique of generating stance explanations outperforms the SOTA extractive summarization method in terms of informativeness, non-redundancy, coverage, and overall quality. Furthermore, experiments show that our explanation-based stance prediction excels or matches the performance of the SOTA model on various benchmark datasets.

Event coreference models cluster event mentions pertaining to the same real-world event. Recent models rely on contextualized representations to recognize coreference among lexically or contextually similar mentions. However, models typically fail to leverage commonsense inferences, which is particularly limiting for resolving lexically-divergent mentions. We propose a model that extends event mentions with temporal commonsense inferences. Given a complex sentence with multiple events, e.g., “the man killed his wife and got arrested”, with the target event “arrested”, our model generates plausible events that happen before the target event – such as “the police arrived”, and after it, such as “he was sentenced”. We show that incorporating such inferences into an existing event coreference model improves its performance, and we analyze the coreferences in which such temporal knowledge is required.

Understanding the speaker’s intended meaning often involves drawing commonsense inferences to reason about what is not stated explicitly. In multi-event sentences, it requires understanding the relationships between events based on contextual knowledge. We propose COMET-M (Multi-Event), an event-centric commonsense model capable of generating commonsense inferences for a target event within a complex sentence. COMET-M builds upon COMET (Bosselut et al., 2019), which excels at generating event-centric inferences for simple sentences, but struggles with the complexity of multi-event sentences prevalent in natural text. To overcome this limitation, we curate a Multi-Event Inference (MEI) dataset of 35K human-written inferences. We train COMET-M on the human-written inferences and also create baselines using automatically labeled examples. Experimental results demonstrate the significant performance improvement of COMET-M over COMET in generating multi-event inferences. Moreover, COMET-M successfully produces distinct inferences for each target event, taking the complete context into consideration. COMET-M holds promise for downstream tasks involving natural text such as coreference resolution, dialogue, and story understanding.

Text classification has wide-ranging applications in various domains. While neural network approaches have drastically advanced performance in text classification, they tend to be powered by a large amount of training data, and interpretability is often an issue. As a step towards better accuracy and interpretability especially on small data, in this paper we present a new knowledge-infused attention mechanism, called KW-ATTN (KnoWledge-infused ATTentioN) to incorporate high-level concepts from external knowledge bases into Neural Network models. We show that KW-ATTN outperforms baseline models using only words as well as other approaches using concepts by classification accuracy, which indicates that high-level concepts help model prediction. Furthermore, crowdsourced human evaluation suggests that additional concept information helps interpretability of the model.

Motivated by applications such as question answering, fact checking, and data integration, there is significant interest in constructing knowledge graphs by extracting information from unstructured information sources, particularly text documents. Knowledge graphs have emerged as a standard for structured knowledge representation, whereby entities and their inter-relations are represented and conveniently stored as (subject,predicate,object) triples in a graph that can be used to power various downstream applications. The proliferation of financial news sources reporting on companies, markets, currencies, and stocks presents an opportunity for extracting valuable knowledge about this crucial domain. In this paper, we focus on constructing a knowledge graph automatically by information extraction from a large corpus of financial news articles. For that purpose, we develop a high precision knowledge extraction pipeline tailored for the financial domain. This pipeline combines multiple information extraction techniques with a financial dictionary that we built, all working together to produce over 342,000 compact extractions from over 288,000 financial news articles, with a precision of 78% at the top-100 extractions. The extracted triples are stored in a knowledge graph making them readily available for use in downstream applications.

Much research has been done within the social sciences on the interpretation and influence of stigma on human behaviour and health, which result in out-of-group exclusion, distancing, cognitive separation, status loss, discrimination, in-group pressure, and often lead to disengagement, non-adherence to treatment plan, and prescriptions by the doctor. However, little work has been conducted on computational identification of stigma in general and in social media discourse in particular. In this paper, we develop the annotation scheme and improve the annotation process for stigma identification, which can be applied to other health-care domains. The data from pro-vaccination and anti-vaccination discussion groups are annotated by trained annotators who have professional background in social science and health-care studies, therefore the group can be considered experts on the subject in comparison to non-expert crowd. Amazon MTurk annotators is another group of annotator with no knowledge on their education background, they are initially treated as non-expert crowd on the subject matter of stigma. We analyze the annotations with visualisation techniques, features from LIWC (Linguistic Inquiry and Word Count) list and make prediction based on bi-grams with traditional and deep learning models. Data augmentation method and application of CNN show high performance accuracy in comparison to other models. Success of the rigorous annotation process on identifying stigma is reconfirmed by achieving high prediction rate with CNN.

Discourse processing is a suite of Natural Language Processing (NLP) tasks to uncover linguistic structures from texts at several levels, which can support many downstream applications. This involves identifying the topic structure, the coherence structure, the coreference structure, and the conversation structure for conversational discourse. Taken together, these structures can inform text summarization, machine translation, essay scoring, sentiment analysis, information extraction, question answering, and thread recovery. The tutorial starts with an overview of basic concepts in discourse analysis – monologue vs. conversation, synchronous vs. asynchronous conversation, and key linguistic structures in discourse analysis. We also give an overview of linguistic structures and corresponding discourse analysis tasks that discourse researchers are generally interested in, as well as key applications on which these discourse structures have an impact.

We define and motivate the problem of summarizing partial email threads. This problem introduces the challenge of generating reference summaries for partial threads when human annotation is only available for the threads as a whole, particularly when the human-selected sentences are not uniformly distributed within the threads. We propose an oracular algorithm for generating these reference summaries with arbitrary length, and we are making the resulting dataset publicly available. In addition, we apply a recent unsupervised method based on Bayesian Surprise that incorporates background knowledge into partial thread summarization, extend it with conversational features, and modify the mechanism by which it handles redundancy. Experiments with our method indicate improved performance over the baseline for shorter partial threads; and our results suggest that the potential benefits of background knowledge to partial thread summarization should be further investigated with larger datasets.

Discourse Parsing and Sentiment Analysis are two fundamental tasks in Natural Language Processing that have been shown to be mutually beneficial. In this work, we design and compare two Neural Based models for jointly learning both tasks. In the proposed approach, we first create a vector representation for all the text segments in the input sentence. Next, we apply three different Recursive Neural Net models: one for discourse structure prediction, one for discourse relation prediction and one for sentiment analysis. Finally, we combine these Neural Nets in two different joint models: Multi-tasking and Pre-training. Our results on two standard corpora indicate that both methods result in improvements in each task but Multi-tasking has a bigger impact than Pre-training. Specifically for Discourse Parsing, we see improvements in the prediction of the set of contrastive relations.

Discourse parsing is a popular technique widely used in text understanding, sentiment analysis and other NLP tasks. However, for most discourse parsers, the performance varies significantly across different discourse relations. In this paper, we first validate the underfitting hypothesis, i.e., the less frequent a relation is in the training data, the poorer the performance on that relation. We then explore how to increase the number of positive training instances, without resorting to manually creating additional labeled data. We propose a training data enrichment framework that relies on co-training of two different discourse parsers on unlabeled documents. Importantly, we show that co-training alone is not sufficient. The framework requires a filtering step to ensure that only “good quality” unlabeled documents can be used for enrichment and re-training. We propose and evaluate two ways to perform the filtering. The first is to use an agreement score between the two parsers. The second is to use only the confidence score of the faster parser. Our empirical results show that agreement score can help to boost the performance on infrequent relations, and that the confidence score is a viable approximation of the agreement score for infrequent relations.