Carolina Scarton

Sebastian T. Vincent, Loïc Barrault, Carolina Scarton · meta-ai

This paper describes the SLT-CDT-UoS group's submission to the first Special Task on Formality Control for Spoken Language Translation, part of the IWSLT 2022 Evaluation Campaign. Our efforts were split between two fronts: data engineering and altering the objective function for best hypothesis selection. We used language-independent methods to extract formal and informal sentence pairs from the provided corpora; using English as a pivot language, we propagated formality annotations to languages treated as zero-shot in the task; we also further improved formality controlling with a hypothesis re-ranking approach. On the test sets for English-to-German and English-to-Spanish, we achieved an average accuracy of .935 within the constrained setting and .995 within unconstrained setting. In a zero-shot setting for English-to-Russian and English-to-Italian, we scored average accuracy of .590 for constrained setting and .659 for unconstrained.

Sebastian T. Vincent, Loïc Barrault, Carolina Scarton · meta-ai

Unlike English, morphologically rich languages can reveal characteristics of speakers or their conversational partners, such as gender and number, via pronouns, morphological endings of words and syntax. When translating from English to such languages, a machine translation model needs to opt for a certain interpretation of textual context, which may lead to serious translation errors if extra-textual information is unavailable. We investigate this challenge in the English-to-Polish language direction. We focus on the underresearched problem of utilising external metadata in automatic translation of TV dialogue, proposing a case study where a wide range of approaches for controlling attributes in translation is employed in a multi-attribute scenario. The best model achieves an improvement of +5.81 chrF++/+6.03 BLEU, with other models achieving competitive performance. We additionally contribute a novel attribute-annotated dataset of Polish TV dialogue and a morphological analysis script used to evaluate attribute control in models.

Tomas Goldsack, Zheheng Luo, Qianqian Xie et al.

This paper presents the results of the shared task on Lay Summarisation of Biomedical Research Articles (BioLaySumm), hosted at the BioNLP Workshop at ACL 2023. The goal of this shared task is to develop abstractive summarisation models capable of generating "lay summaries" (i.e., summaries that are comprehensible to non-technical audiences) in both a controllable and non-controllable setting. There are two subtasks: 1) Lay Summarisation, where the goal is for participants to build models for lay summary generation only, given the full article text and the corresponding abstract as input; and 2) Readability-controlled Summarisation, where the goal is for participants to train models to generate both the technical abstract and the lay summary, given an article's main text as input. In addition to overall results, we report on the setup and insights from the BioLaySumm shared task, which attracted a total of 20 participating teams across both subtasks.

Tomas Goldsack, Zhihao Zhang, Chenghua Lin et al.

Lay summarisation aims to jointly summarise and simplify a given text, thus making its content more comprehensible to non-experts. Automatic approaches for lay summarisation can provide significant value in broadening access to scientific literature, enabling a greater degree of both interdisciplinary knowledge sharing and public understanding when it comes to research findings. However, current corpora for this task are limited in their size and scope, hindering the development of broadly applicable data-driven approaches. Aiming to rectify these issues, we present two novel lay summarisation datasets, PLOS (large-scale) and eLife (medium-scale), each of which contains biomedical journal articles alongside expert-written lay summaries. We provide a thorough characterisation of our lay summaries, highlighting differing levels of readability and abstractiveness between datasets that can be leveraged to support the needs of different applications. Finally, we benchmark our datasets using mainstream summarisation approaches and perform a manual evaluation with domain experts, demonstrating their utility and casting light on the key challenges of this task.

Harish Tayyar Madabushi, Edward Gow-Smith, Marcos Garcia et al.

This paper presents the shared task on Multilingual Idiomaticity Detection and Sentence Embedding, which consists of two subtasks: (a) a binary classification task aimed at identifying whether a sentence contains an idiomatic expression, and (b) a task based on semantic text similarity which requires the model to adequately represent potentially idiomatic expressions in context. Each subtask includes different settings regarding the amount of training data. Besides the task description, this paper introduces the datasets in English, Portuguese, and Galician and their annotation procedure, the evaluation metrics, and a summary of the participant systems and their results. The task had close to 100 registered participants organised into twenty five teams making over 650 and 150 submissions in the practice and evaluation phases respectively.

Olesya Razuvayevskaya, Ben Wu, Joao A. Leite et al.

Adapters and Low-Rank Adaptation (LoRA) are parameter-efficient fine-tuning techniques designed to make the training of language models more efficient. Previous results demonstrated that these methods can even improve performance on some classification tasks. This paper complements the existing research by investigating how these techniques influence the classification performance and computation costs compared to full fine-tuning when applied to multilingual text classification tasks (genre, framing, and persuasion techniques detection; with different input lengths, number of predicted classes and classification difficulty), some of which have limited training data. In addition, we conduct in-depth analyses of their efficacy across different training scenarios (training on the original multilingual data; on the translations into English; and on a subset of English-only data) and different languages. Our findings provide valuable insights into the applicability of the parameter-efficient fine-tuning techniques, particularly to complex multilingual and multilabel classification tasks.

Yida Mu, Mali Jin, Charlie Grimshaw et al.

Vaccine hesitancy has been a common concern, probably since vaccines were created and, with the popularisation of social media, people started to express their concerns about vaccines online alongside those posting pro- and anti-vaccine content. Predictably, since the first mentions of a COVID-19 vaccine, social media users posted about their fears and concerns or about their support and belief into the effectiveness of these rapidly developing vaccines. Identifying and understanding the reasons behind public hesitancy towards COVID-19 vaccines is important for policy markers that need to develop actions to better inform the population with the aim of increasing vaccine take-up. In the case of COVID-19, where the fast development of the vaccines was mirrored closely by growth in anti-vaxx disinformation, automatic means of detecting citizen attitudes towards vaccination became necessary. This is an important computational social sciences task that requires data analysis in order to gain in-depth understanding of the phenomena at hand. Annotated data is also necessary for training data-driven models for more nuanced analysis of attitudes towards vaccination. To this end, we created a new collection of over 3,101 tweets annotated with users' attitudes towards COVID-19 vaccination (stance). Besides, we also develop a domain-specific language model (VaxxBERT) that achieves the best predictive performance (73.0 accuracy and 69.3 F1-score) as compared to a robust set of baselines. To the best of our knowledge, these are the first dataset and model that model vaccine hesitancy as a category distinct from pro- and anti-vaccine stance.

Tomas Goldsack, Carolina Scarton, Matthew Shardlow et al.

This paper presents the setup and results of the second edition of the BioLaySumm shared task on the Lay Summarisation of Biomedical Research Articles, hosted at the BioNLP Workshop at ACL 2024. In this task edition, we aim to build on the first edition's success by further increasing research interest in this important task and encouraging participants to explore novel approaches that will help advance the state-of-the-art. Encouragingly, we found research interest in the task to be high, with this edition of the task attracting a total of 53 participating teams, a significant increase in engagement from the previous edition. Overall, our results show that a broad range of innovative approaches were adopted by task participants, with a predictable shift towards the use of Large Language Models (LLMs).

Edward Gow-Smith, Harish Tayyar Madabushi, Carolina Scarton et al.

Tokenisation is the first step in almost all NLP tasks, and state-of-the-art transformer-based language models all use subword tokenisation algorithms to process input text. Existing algorithms have problems, often producing tokenisations of limited linguistic validity, and representing equivalent strings differently depending on their position within a word. We hypothesise that these problems hinder the ability of transformer-based models to handle complex words, and suggest that these problems are a result of allowing tokens to include spaces. We thus experiment with an alternative tokenisation approach where spaces are always treated as individual tokens. Specifically, we apply this modification to the BPE and Unigram algorithms. We find that our modified algorithms lead to improved performance on downstream NLP tasks that involve handling complex words, whilst having no detrimental effect on performance in general natural language understanding tasks. Intrinsically, we find our modified algorithms give more morphologically correct tokenisations, in particular when handling prefixes. Given the results of our experiments, we advocate for always treating spaces as individual tokens as an improved tokenisation method.

Ben Wu, Olesya Razuvayevskaya, Freddy Heppell et al.

This paper describes our approach for SemEval-2023 Task 3: Detecting the category, the framing, and the persuasion techniques in online news in a multi-lingual setup. For Subtask 1 (News Genre), we propose an ensemble of fully trained and adapter mBERT models which was ranked joint-first for German, and had the highest mean rank of multi-language teams. For Subtask 2 (Framing), we achieved first place in 3 languages, and the best average rank across all the languages, by using two separate ensembles: a monolingual RoBERTa-MUPPETLARGE and an ensemble of XLM-RoBERTaLARGE with adapters and task adaptive pretraining. For Subtask 3 (Persuasion Techniques), we train a monolingual RoBERTa-Base model for English and a multilingual mBERT model for the remaining languages, which achieved top 10 for all languages, including 2nd for English. For each subtask, we compared monolingual and multilingual approaches, and considered class imbalance techniques.

João A. Leite, Olesya Razuvayevskaya, Kalina Bontcheva et al.

Credibility signals represent a wide range of heuristics typically used by journalists and fact-checkers to assess the veracity of online content. Automating the extraction of credibility signals presents significant challenges due to the necessity of training high-accuracy, signal-specific extractors, coupled with the lack of sufficiently large annotated datasets. This paper introduces Pastel (Prompted weAk Supervision wiTh crEdibility signaLs), a weakly supervised approach that leverages large language models (LLMs) to extract credibility signals from web content, and subsequently combines them to predict the veracity of content without relying on human supervision. We validate our approach using four article-level misinformation detection datasets, demonstrating that Pastel outperforms zero-shot veracity detection by 38.3% and achieves 86.7% of the performance of the state-of-the-art system trained with human supervision. Moreover, in cross-domain settings where training and testing datasets originate from different domains, Pastel significantly outperforms the state-of-the-art supervised model by 63%. We further study the association between credibility signals and veracity, and perform an ablation study showing the impact of each signal on model performance. Our findings reveal that 12 out of the 19 proposed signals exhibit strong associations with veracity across all datasets, while some signals show domain-specific strengths.

Iknoor Singh, Carolina Scarton, Xingyi Song et al.

Finding previously debunked narratives involves identifying claims that have already undergone fact-checking. The issue intensifies when similar false claims persist in multiple languages, despite the availability of debunks for several months in another language. Hence, automatically finding debunks (or fact-checks) in multiple languages is crucial to make the best use of scarce fact-checkers' resources. Mainly due to the lack of readily available data, this is an understudied problem, particularly when considering the cross-lingual scenario, i.e. the retrieval of debunks in a language different from the language of the online post being checked. This study introduces cross-lingual debunked narrative retrieval and addresses this research gap by: (i) creating Multilingual Misinformation Tweets (MMTweets): a dataset that stands out, featuring cross-lingual pairs, images, human annotations, and fine-grained labels, making it a comprehensive resource compared to its counterparts; (ii) conducting an extensive experiment to benchmark state-of-the-art cross-lingual retrieval models and introducing multistage retrieval methods tailored for the task; and (iii) comprehensively evaluating retrieval models for their cross-lingual and cross-dataset transfer capabilities within MMTweets, and conducting a retrieval latency analysis. We find that MMTweets presents challenges for cross-lingual debunked narrative retrieval, highlighting areas for improvement in retrieval models. Nonetheless, the study provides valuable insights for creating MMTweets datasets and optimising debunked narrative retrieval models to empower fact-checking endeavours. The dataset and annotation codebook are publicly available at https://doi.org/10.5281/zenodo.10637161.

Sebastian Vincent, Alice Dowek, Rowanne Sumner et al.

Sensitising language models (LMs) to external context helps them to more effectively capture the speaking patterns of individuals with specific characteristics or in particular environments. This work investigates to what extent rich character and film annotations can be leveraged to personalise LMs in a scalable manner. We then explore the use of such models in evaluating context specificity in machine translation. We build LMs which leverage rich contextual information to reduce perplexity by up to 6.5% compared to a non-contextual model, and generalise well to a scenario with no speaker-specific data, relying on combinations of demographic characteristics expressed via metadata. Our findings are consistent across two corpora, one of which (Cornell-rich) is also a contribution of this paper. We then use our personalised LMs to measure the co-occurrence of extra-textual context and translation hypotheses in a machine translation setting. Our results suggest that the degree to which professional translations in our domain are context-specific can be preserved to a better extent by a contextual machine translation model than a non-contextual model, which is also reflected in the contextual model's superior reference-based scores.

Ben Wu, Yue Li, Yida Mu et al.

In this paper, we address the limitations of the common data annotation and training methods for objective single-label classification tasks. Typically, when annotating such tasks annotators are only asked to provide a single label for each sample and annotator disagreement is discarded when a final hard label is decided through majority voting. We challenge this traditional approach, acknowledging that determining the appropriate label can be difficult due to the ambiguity and lack of context in the data samples. Rather than discarding the information from such ambiguous annotations, our soft label method makes use of them for training. Our findings indicate that additional annotator information, such as confidence, secondary label and disagreement, can be used to effectively generate soft labels. Training classifiers with these soft labels then leads to improved performance and calibration on the hard label test set.

Tomas Goldsack, Zhihao Zhang, Chen Tang et al.

Previous approaches for automatic lay summarisation are exclusively reliant on the source article that, given it is written for a technical audience (e.g., researchers), is unlikely to explicitly define all technical concepts or state all of the background information that is relevant for a lay audience. We address this issue by augmenting eLife, an existing biomedical lay summarisation dataset, with article-specific knowledge graphs, each containing detailed information on relevant biomedical concepts. Using both automatic and human evaluations, we systematically investigate the effectiveness of three different approaches for incorporating knowledge graphs within lay summarisation models, with each method targeting a distinct area of the encoder-decoder model architecture. Our results confirm that integrating graph-based domain knowledge can significantly benefit lay summarisation by substantially increasing the readability of generated text and improving the explanation of technical concepts.

Dylan Phelps, Xuan-Rui Fan, Edward Gow-Smith et al.

Deep neural models, in particular Transformer-based pre-trained language models, require a significant amount of data to train. This need for data tends to lead to problems when dealing with idiomatic multiword expressions (MWEs), which are inherently less frequent in natural text. As such, this work explores sample efficient methods of idiomaticity detection. In particular we study the impact of Pattern Exploit Training (PET), a few-shot method of classification, and BERTRAM, an efficient method of creating contextual embeddings, on the task of idiomaticity detection. In addition, to further explore generalisability, we focus on the identification of MWEs not present in the training data. Our experiments show that while these methods improve performance on English, they are much less effective on Portuguese and Galician, leading to an overall performance about on par with vanilla mBERT. Regardless, we believe sample efficient methods for both identifying and representing potentially idiomatic MWEs are very encouraging and hold significant potential for future exploration.

Yida Mu, Ye Jiang, Freddy Heppell et al.

The COVID-19 pandemic led to an infodemic where an overwhelming amount of COVID-19 related content was being disseminated at high velocity through social media. This made it challenging for citizens to differentiate between accurate and inaccurate information about COVID-19. This motivated us to carry out a comparative study of the characteristics of COVID-19 misinformation versus those of accurate COVID-19 information through a large-scale computational analysis of over 242 million tweets. The study makes comparisons alongside four key aspects: 1) the distribution of topics, 2) the live status of tweets, 3) language analysis and 4) the spreading power over time. An added contribution of this study is the creation of a COVID-19 misinformation classification dataset. Finally, we demonstrate that this new dataset helps improve misinformation classification by more than 9\% based on average F1 measure.

Yue Li, Carolina Scarton, Xingyi Song et al.

Vaccine hesitancy is widespread, despite the government's information campaigns and the efforts of the World Health Organisation (WHO). Categorising the topics within vaccine-related narratives is crucial to understand the concerns expressed in discussions and identify the specific issues that contribute to vaccine hesitancy. This paper addresses the need for monitoring and analysing vaccine narratives online by introducing a novel vaccine narrative classification task, which categorises COVID-19 vaccine claims into one of seven categories. Following a data augmentation approach, we first construct a novel dataset for this new classification task, focusing on the minority classes. We also make use of fact-checker annotated data. The paper also presents a neural vaccine narrative classifier that achieves an accuracy of 84% under cross-validation. The classifier is publicly available for researchers and journalists.

Freddy Heppell, Kalina Bontcheva, Carolina Scarton

This paper analyses two hitherto unstudied sites sharing state-backed disinformation, Reliable Recent News (rrn.world) and WarOnFakes (waronfakes.com), which publish content in Arabic, Chinese, English, French, German, and Spanish. We describe our content acquisition methodology and perform cross-site unsupervised topic clustering on the resulting multilingual dataset. We also perform linguistic and temporal analysis of the web page translations and topics over time, and investigate articles with false publication dates. We make publicly available this new dataset of 14,053 articles, annotated with each language version, and additional metadata such as links and images. The main contribution of this paper for the NLP community is in the novel dataset which enables studies of disinformation networks, and the training of NLP tools for disinformation detection.

Iknoor Singh, Yue Li, Melissa Thong et al.

This paper describes the second-placed system on the leaderboard of SemEval-2022 Task 8: Multilingual News Article Similarity. We propose an entity-enriched Siamese Transformer which computes news article similarity based on different sub-dimensions, such as the shared narrative, entities, location and time of the event discussed in the news article. Our system exploits a Siamese network architecture using a Transformer encoder to learn document-level representations for the purpose of capturing the narrative together with the auxiliary entity-based features extracted from the news articles. The intuition behind using all these features together is to capture the similarity between news articles at different granularity levels and to assess the extent to which different news outlets write about "the same events". Our experimental results and detailed ablation study demonstrate the effectiveness and the validity of our proposed method.

João A. Leite, Carolina Scarton, Diego F. Silva

Online social media is rife with offensive and hateful comments, prompting the need for their automatic detection given the sheer amount of posts created every second. Creating high-quality human-labelled datasets for this task is difficult and costly, especially because non-offensive posts are significantly more frequent than offensive ones. However, unlabelled data is abundant, easier, and cheaper to obtain. In this scenario, self-training methods, using weakly-labelled examples to increase the amount of training data, can be employed. Recent "noisy" self-training approaches incorporate data augmentation techniques to ensure prediction consistency and increase robustness against noisy data and adversarial attacks. In this paper, we experiment with default and noisy self-training using three different textual data augmentation techniques across five different pre-trained BERT architectures varying in size. We evaluate our experiments on two offensive/hate-speech datasets and demonstrate that (i) self-training consistently improves performance regardless of model size, resulting in up to +1.5% F1-macro on both datasets, and (ii) noisy self-training with textual data augmentations, despite being successfully applied in similar settings, decreases performance on offensive and hate-speech domains when compared to the default method, even with state-of-the-art augmentations such as backtranslation.

Yue Li, Carolina Scarton

Considering a conversation thread, rumour stance classification aims to identify the opinion (e.g. agree or disagree) of replies towards a target (rumour story). Although the target is expected to be an essential component in traditional stance classification, we show that rumour stance classification datasets contain a considerable amount of real-world data whose stance could be naturally inferred directly from the replies, contributing to the strong performance of the supervised models without awareness of the target. We find that current target-aware models underperform in cases where the context of the target is crucial. Finally, we propose a simple yet effective framework to enhance reasoning with the targets, achieving state-of-the-art performance on two benchmark datasets.

João A. Leite, Olesya Razuvayevskaya, Kalina Bontcheva et al.

Automated fact-checking (AFC) systems are susceptible to adversarial attacks, enabling false claims to evade detection. Existing adversarial frameworks typically rely on injecting noise or altering semantics, yet no existing framework exploits the adversarial potential of persuasion techniques, which are widely used in disinformation campaigns to manipulate audiences. In this paper, we introduce a novel class of persuasive adversarial attacks on AFCs by employing a generative LLM to rephrase claims using persuasion techniques. Considering 15 techniques grouped into 6 categories, we study the effects of persuasion on both claim verification and evidence retrieval using a decoupled evaluation strategy. Experiments on the FEVER and FEVEROUS benchmarks show that persuasion attacks can substantially degrade both verification performance and evidence retrieval. Our analysis identifies persuasion techniques as a potent class of adversarial attacks, highlighting the need for more robust AFC systems.

Iknoor Singh, Carolina Scarton, Kalina Bontcheva

The proliferation of online news and the increasing spread of misinformation necessitate robust methods for automatic data analysis. Narrative classification is emerging as a important task, since identifying what is being said online is critical for fact-checkers, policy markers and other professionals working on information studies. This paper presents our approach to SemEval 2025 Task 10 Subtask 2, which aims to classify news articles into a pre-defined two-level taxonomy of main narratives and sub-narratives across multiple languages. We propose Hierarchical Three-Step Prompting (H3Prompt) for multilingual narrative classification. Our methodology follows a three-step Large Language Model (LLM) prompting strategy, where the model first categorises an article into one of two domains (Ukraine-Russia War or Climate Change), then identifies the most relevant main narratives, and finally assigns sub-narratives. Our approach secured the top position on the English test set among 28 competing teams worldwide. The code is available at https://github.com/GateNLP/H3Prompt.

João A. Leite, Olesya Razuvayevskaya, Carolina Scarton et al.

Disinformation, irrespective of domain or language, aims to deceive or manipulate public opinion, typically through employing advanced persuasion techniques. Qualitative and quantitative research on the weaponisation of persuasion techniques in disinformation has been mostly topic-specific (e.g., COVID-19) with limited cross-domain studies, resulting in a lack of comprehensive understanding of these strategies. This study employs a state-of-the-art persuasion technique classifier to conduct a large-scale, multi-domain analysis of the role of 16 persuasion techniques in disinformation narratives. It shows how different persuasion techniques are employed disproportionately in different disinformation domains. We also include a detailed case study on climate change disinformation, highlighting how linguistic, psychological, and cultural factors shape the adaptation of persuasion strategies to fit unique thematic contexts.

Jake Vasilakes, Carolina Scarton, Zhixue Zhao

Social media's global reach amplifies the spread of information, highlighting the need for robust Natural Language Processing tasks like stance detection across languages and modalities. Prior research predominantly focuses on text-only inputs, leaving multimodal scenarios, such as those involving both images and text, relatively underexplored. Meanwhile, the prevalence of multimodal posts has increased significantly in recent years. Although state-of-the-art Vision-Language Models (VLMs) show promise, their performance on multimodal and multilingual stance detection tasks remains largely unexamined. This paper evaluates state-of-the-art VLMs on a newly extended dataset covering seven languages and multimodal inputs, investigating their use of visual cues, language-specific performance, and cross-modality interactions. Our results show that VLMs generally rely more on text than images for stance detection and this trend persists across languages. Additionally, VLMs rely significantly more on text contained within the images than other visual content. Regarding multilinguality, the models studied tend to generate consistent predictions across languages whether they are explicitly multilingual or not, although there are outliers that are incongruous with macro F1, language support, and model size.

Tomas Goldsack, Carolina Scarton, Chenghua Lin

In this work, we explore the application of Large Language Models to zero-shot Lay Summarisation. We propose a novel two-stage framework for Lay Summarisation based on real-life processes, and find that summaries generated with this method are increasingly preferred by human judges for larger models. To help establish best practices for employing LLMs in zero-shot settings, we also assess the ability of LLMs as judges, finding that they are able to replicate the preferences of human judges. Finally, we take the initial steps towards Lay Summarisation for Natural Language Processing (NLP) articles, finding that LLMs are able to generalise to this new domain, and further highlighting the greater utility of summaries generated by our proposed approach via an in-depth human evaluation.

Ivan Srba, Olesya Razuvayevskaya, João A. Leite et al.

In the age of social media and generative AI, the ability to automatically assess the credibility of online content has become increasingly critical, complementing traditional approaches to false information detection. Credibility assessment relies on aggregating diverse credibility signals - small units of information, such as content subjectivity, bias, or a presence of persuasion techniques - into a final credibility label/score. However, current research in automatic credibility assessment and credibility signals detection remains highly fragmented, with many signals studied in isolation and lacking integration. Notably, there is a scarcity of approaches that detect and aggregate multiple credibility signals simultaneously. These challenges are further exacerbated by the absence of a comprehensive and up-to-date overview of research works that connects these research efforts under a common framework and identifies shared trends, challenges, and open problems. In this survey, we address this gap by presenting a systematic and comprehensive literature review of 175 research papers, focusing on textual credibility signals within the field of Natural Language Processing (NLP), which undergoes a rapid transformation due to advancements in Large Language Models (LLMs). While positioning the NLP research into the the broader multidisciplinary landscape, we examine both automatic credibility assessment methods as well as the detection of nine categories of credibility signals. We provide an in-depth analysis of three key categories: 1) factuality, subjectivity and bias, 2) persuasion techniques and logical fallacies, and 3) check-worthy and fact-checked claims. In addition to summarising existing methods, datasets, and tools, we outline future research direction and emerging opportunities, with particular attention to evolving challenges posed by generative AI.

Edward Gow-Smith, Dylan Phelps, Harish Tayyar Madabushi et al.

All existing transformer-based approaches to NLP using subword tokenisation algorithms encode whitespace (word boundary information) through the use of special space symbols (such as \#\# or \_) forming part of tokens. These symbols have been shown to a) lead to reduced morphological validity of tokenisations, and b) give substantial vocabulary redundancy. As such, removing these symbols has been shown to have a beneficial effect on the processing of morphologically complex words for transformer encoders in the pretrain-finetune paradigm. In this work, we explore whether word boundary information is at all useful to such models. In particular, we train transformer encoders across four different training scales, and investigate several alternative approaches to including word boundary information, evaluating on a range of tasks across different domains and problem set-ups: GLUE (for sentence-level classification), NER (for token-level classification), and two classification datasets involving complex words (Superbizarre and FLOTA). Overall, through an extensive experimental setup that includes the pre-training of 29 models, we find no substantial improvements from our alternative approaches, suggesting that modifying tokenisers to remove word boundary information isn't leading to a loss of useful information.

Yue Li, Zhixue Zhao, Carolina Scarton

Extremely low-resource languages, especially those written in rare scripts, as shown in Figure 1, remain largely unsupported by large language models (LLMs). This is due in part to compounding factors such as the lack of training data. This paper delivers the first comprehensive analysis of whether LLMs can acquire such languages purely via in-context learning (ICL), with or without auxiliary alignment signals, and how these methods compare to parameter-efficient fine-tuning (PEFT). We systematically evaluate 20 under-represented languages across three state-of-the-art multilingual LLMs. Our findings highlight the limitation of PEFT when both language and its script are extremely under-represented by the LLM. In contrast, zero-shot ICL with language alignment is impressively effective on extremely low-resource languages, while few-shot ICL or PEFT is more beneficial for languages relatively better represented by LLMs. For LLM practitioners working on extremely low-resource languages, we summarise guidelines grounded by our results on adapting LLMs to low-resource languages, e.g., avoiding fine-tuning a multilingual model on languages of unseen scripts.

Wei He, Tiago Kramer Vieira, Marcos Garcia et al.

Idiomatic expressions are an integral part of human languages, often used to express complex ideas in compressed or conventional ways (e.g. eager beaver as a keen and enthusiastic person). However, their interpretations may not be straightforwardly linked to the meanings of their individual components in isolation and this may have an impact for compositional approaches. In this paper, we investigate to what extent word representation models are able to go beyond compositional word combinations and capture multiword expression idiomaticity and some of the expected properties related to idiomatic meanings. We focus on noun compounds of varying levels of idiomaticity in two languages (English and Portuguese), presenting a dataset of minimal pairs containing human idiomaticity judgments for each noun compound at both type and token levels, their paraphrases and their occurrences in naturalistic and sense-neutral contexts, totalling 32,200 sentences. We propose this set of minimal pairs for evaluating how well a model captures idiomatic meanings, and define a set of fine-grained metrics of Affinity and Scaled Similarity, to determine how sensitive the models are to perturbations that may lead to changes in idiomaticity. The results obtained with a variety of representative and widely used models indicate that, despite superficial indications to the contrary in the form of high similarities, idiomaticity is not yet accurately represented in current models. Moreover, the performance of models with different levels of contextualisation suggests that their ability to capture context is not yet able to go beyond more superficial lexical clues provided by the words and to actually incorporate the relevant semantic clues needed for idiomaticity.

Yue Li, Zhixue Zhao, Carolina Scarton

In-context learning (ICL) performance is highly sensitive to prompt design, yet the impact of class label options (e.g. lexicon or order) in zero-shot classification remains underexplored. This study proposes LOADS (Label set Optimization via Activation Distribution kurtosiS), a post-hoc method for selecting optimal label sets in zero-shot ICL with large language models (LLMs). LOADS is built upon the observations in our empirical analysis, the first to systematically examine how label option design (i.e., lexical choice, order, and elaboration) impacts classification performance. This analysis shows that the lexical choice of the labels in the prompt (such as agree vs. support in stance classification) plays an important role in both model performance and model's sensitivity to the label order. A further investigation demonstrates that optimal label words tend to activate fewer outlier neurons in LLMs' feed-forward networks. LOADS then leverages kurtosis to measure the neuron activation distribution for label selection, requiring only a single forward pass without gradient propagation or labelled data. The LOADS-selected label words consistently demonstrate effectiveness for zero-shot ICL across classification tasks, datasets, models and languages, achieving maximum performance gain from 0.54 to 0.76 compared to the conventional approach of using original dataset label words.

João A. Leite, Arnav Arora, Silvia Gargova et al.

Large Language Models (LLMs) can generate human-like disinformation, yet their ability to personalise such content across languages and demographics remains underexplored. This study presents the first large-scale, multilingual analysis of persona-targeted disinformation generation by LLMs. Employing a red teaming methodology, we prompt eight state-of-the-art LLMs with 324 false narratives and 150 demographic personas (combinations of country, generation, and political orientation) across four languages--English, Russian, Portuguese, and Hindi--resulting in AI-TRAITS, a comprehensive dataset of 1.6 million personalised disinformation texts. Results show that the use of even simple personalisation prompts significantly increases the likelihood of jailbreaks across all studied LLMs, up to 10 percentage points, and alters linguistic and rhetorical patterns that enhance narrative persuasiveness. Models such as Grok and GPT exhibited jailbreak rates and personalisation scores both exceeding 85%. These insights expose critical vulnerabilities in current state-of-the-art LLMs and offer a foundation for improving safety alignment and detection strategies in multilingual and cross-demographic contexts.

Yue Li, Jake Vasilakes, Zhixue Zhao et al.

We introduce SCRum-9, the largest multilingual Stance Classification dataset for Rumour analysis in 9 languages, containing 7,516 tweets from X. SCRum-9 goes beyond existing stance classification datasets by covering more languages, linking examples to more fact-checked claims (2.1k), and including confidence-related annotations from multiple annotators to account for intra- and inter-annotator variability. Annotations were made by at least two native speakers per language, totalling more than 405 hours of annotation and 8,150 dollars in compensation. Further, SCRum-9 is used to benchmark five large language models (LLMs) and two multilingual masked language models (MLMs) in In-Context Learning (ICL) and fine-tuning setups. This paper also innovates by exploring the use of multilingual synthetic data for rumour stance classification, showing that even LLMs with weak ICL performance can produce valuable synthetic data for fine-tuning small MLMs, enabling them to achieve higher performance than zero-shot ICL in LLMs. Finally, we examine the relationship between model predictions and human uncertainty on ambiguous cases finding that model predictions often match the second-choice labels assigned by annotators, rather than diverging entirely from human judgments. SCRum-9 is publicly released to the research community with potential to foster further research on multilingual analysis of misleading narratives on social media.

Fatima Haouari, Carolina Scarton, Nicolò Faggiani et al.

Misleading narratives play a crucial role in shaping public opinion during elections, as they can influence how voters perceive candidates and political parties. This entails the need to detect these narratives accurately. To address this, we introduce the first taxonomy of common misleading narratives that circulated during recent elections in Europe. Based on this taxonomy, we construct and analyse UKElectionNarratives: the first dataset of human-annotated misleading narratives which circulated during the UK General Elections in 2019 and 2024. We also benchmark Pre-trained and Large Language Models (focusing on GPT-4o), studying their effectiveness in detecting election-related misleading narratives. Finally, we discuss potential use cases and make recommendations for future research directions using the proposed codebook and dataset.

Sebastian Vincent, Charlotte Prescott, Chris Bayliss et al.

Incorporating extra-textual context such as film metadata into the machine translation (MT) pipeline can enhance translation quality, as indicated by automatic evaluation in recent work. However, the positive impact of such systems in industry remains unproven. We report on an industrial case study carried out to investigate the benefit of MT in a professional scenario of translating TV subtitles with a focus on how leveraging extra-textual context impacts post-editing. We found that post-editors marked significantly fewer context-related errors when correcting the outputs of MTCue, the context-aware model, as opposed to non-contextual models. We also present the results of a survey of the employed post-editors, which highlights contextual inadequacy as a significant gap consistently observed in MT. Our findings strengthen the motivation for further work within fully contextual MT.

Wei He, Marco Idiart, Carolina Scarton et al.

Accurately modeling idiomatic or non-compositional language has been a longstanding challenge in Natural Language Processing (NLP). This is partly because these expressions do not derive their meanings solely from their constituent words, but also due to the scarcity of relevant data resources, and their impact on the performance of downstream tasks such as machine translation and simplification. In this paper we propose an approach to model idiomaticity effectively using a triplet loss that incorporates the asymmetric contribution of components words to an idiomatic meaning for training language models by using adaptive contrastive learning and resampling miners to build an idiomatic-aware learning objective. Our proposed method is evaluated on a SemEval challenge and outperforms previous alternatives significantly in many metrics.

João A. Leite, Olesya Razuvayevskaya, Kalina Bontcheva et al.

This work introduces EUvsDisinfo, a multilingual dataset of disinformation articles originating from pro-Kremlin outlets, along with trustworthy articles from credible / less biased sources. It is sourced directly from the debunk articles written by experts leading the EUvsDisinfo project. Our dataset is the largest to-date resource in terms of the overall number of articles and distinct languages. It also provides the largest topical and temporal coverage. Using this dataset, we investigate the dissemination of pro-Kremlin disinformation across different languages, uncovering language-specific patterns targeting certain disinformation topics. We further analyse the evolution of topic distribution over an eight-year period, noting a significant surge in disinformation content before the full-scale invasion of Ukraine in 2022. Lastly, we demonstrate the dataset's applicability in training models to effectively distinguish between disinformation and trustworthy content in multilingual settings.

Zhihao Zhang, Tomas Goldsack, Carolina Scarton et al.

Lay summarisation aims to produce summaries of scientific articles that are comprehensible to non-expert audiences. However, previous work assumes a one-size-fits-all approach, where the content and style of the produced summary are entirely dependent on the data used to train the model. In practice, audiences with different levels of expertise will have specific needs, impacting what content should appear in a lay summary and how it should be presented. Aiming to address this, we propose ATLAS, a novel abstractive summarisation approach that can control various properties that contribute to the overall "layness" of the generated summary using targeted control attributes. We evaluate ATLAS on a combination of biomedical lay summarisation datasets, where it outperforms state-of-the-art baselines using mainstream summarisation metrics. Additional analyses provided on the discriminatory power and emergent influence of our selected controllable attributes further attest to the effectiveness of our approach.

Sebastian Vincent, Robert Flynn, Carolina Scarton

Efficient utilisation of both intra- and extra-textual context remains one of the critical gaps between machine and human translation. Existing research has primarily focused on providing individual, well-defined types of context in translation, such as the surrounding text or discrete external variables like the speaker's gender. This work introduces MTCue, a novel neural machine translation (NMT) framework that interprets all context (including discrete variables) as text. MTCue learns an abstract representation of context, enabling transferability across different data settings and leveraging similar attributes in low-resource scenarios. With a focus on a dialogue domain with access to document and metadata context, we extensively evaluate MTCue in four language pairs in both translation directions. Our framework demonstrates significant improvements in translation quality over a parameter-matched non-contextual baseline, as measured by BLEU (+0.88) and Comet (+1.58). Moreover, MTCue significantly outperforms a "tagging" baseline at translating English text. Analysis reveals that the context encoder of MTCue learns a representation space that organises context based on specific attributes, such as formality, enabling effective zero-shot control. Pre-training on context embeddings also improves MTCue's few-shot performance compared to the "tagging" baseline. Finally, an ablation study conducted on model components and contextual variables further supports the robustness of MTCue for context-based NMT.

Yida Mu, Ben P. Wu, William Thorne et al.

Instruction-tuned Large Language Models (LLMs) have exhibited impressive language understanding and the capacity to generate responses that follow specific prompts. However, due to the computational demands associated with training these models, their applications often adopt a zero-shot setting. In this paper, we evaluate the zero-shot performance of two publicly accessible LLMs, ChatGPT and OpenAssistant, in the context of six Computational Social Science classification tasks, while also investigating the effects of various prompting strategies. Our experiments investigate the impact of prompt complexity, including the effect of incorporating label definitions into the prompt; use of synonyms for label names; and the influence of integrating past memories during foundation model training. The findings indicate that in a zero-shot setting, current LLMs are unable to match the performance of smaller, fine-tuned baseline transformer models (such as BERT-large). Additionally, we find that different prompting strategies can significantly affect classification accuracy, with variations in accuracy and F1 scores exceeding 10\%.

Harish Tayyar Madabushi, Edward Gow-Smith, Carolina Scarton et al.

Despite their success in a variety of NLP tasks, pre-trained language models, due to their heavy reliance on compositionality, fail in effectively capturing the meanings of multiword expressions (MWEs), especially idioms. Therefore, datasets and methods to improve the representation of MWEs are urgently needed. Existing datasets are limited to providing the degree of idiomaticity of expressions along with the literal and, where applicable, (a single) non-literal interpretation of MWEs. This work presents a novel dataset of naturally occurring sentences containing MWEs manually classified into a fine-grained set of meanings, spanning both English and Portuguese. We use this dataset in two tasks designed to test i) a language model's ability to detect idiom usage, and ii) the effectiveness of a language model in generating representations of sentences containing idioms. Our experiments demonstrate that, on the task of detecting idiomatic usage, these models perform reasonably well in the one-shot and few-shot scenarios, but that there is significant scope for improvement in the zero-shot scenario. On the task of representing idiomaticity, we find that pre-training is not always effective, while fine-tuning could provide a sample efficient method of learning representations of sentences containing MWEs.

Ye Jiang, Xingyi Song, Carolina Scarton et al.

The spreading COVID-19 misinformation over social media already draws the attention of many researchers. According to Google Scholar, about 26000 COVID-19 related misinformation studies have been published to date. Most of these studies focusing on 1) detect and/or 2) analysing the characteristics of COVID-19 related misinformation. However, the study of the social behaviours related to misinformation is often neglected. In this paper, we introduce a fine-grained annotated misinformation tweets dataset including social behaviours annotation (e.g. comment or question to the misinformation). The dataset not only allows social behaviours analysis but also suitable for both evidence-based or non-evidence-based misinformation classification task. In addition, we introduce leave claim out validation in our experiments and demonstrate the misinformation classification performance could be significantly different when applying to real-world unseen misinformation.

Iknoor Singh, Carolina Scarton, Kalina Bontcheva

The Coronavirus (COVID-19) pandemic has led to a rapidly growing 'infodemic' of health information online. This has motivated the need for accurate semantic search and retrieval of reliable COVID-19 information across millions of documents, in multiple languages. To address this challenge, this paper proposes a novel high precision and high recall neural Multistage BiCross encoder approach. It is a sequential three-stage ranking pipeline which uses the Okapi BM25 retrieval algorithm and transformer-based bi-encoder and cross-encoder to effectively rank the documents with respect to the given query. We present experimental results from our participation in the Multilingual Information Access (MLIA) shared task on COVID-19 multilingual semantic search. The independently evaluated MLIA results validate our approach and demonstrate that it outperforms other state-of-the-art approaches according to nearly all evaluation metrics in cases of both monolingual and bilingual runs.

João A. Leite, Diego F. Silva, Kalina Bontcheva et al.

Hate speech and toxic comments are a common concern of social media platform users. Although these comments are, fortunately, the minority in these platforms, they are still capable of causing harm. Therefore, identifying these comments is an important task for studying and preventing the proliferation of toxicity in social media. Previous work in automatically detecting toxic comments focus mainly in English, with very few work in languages like Brazilian Portuguese. In this paper, we propose a new large-scale dataset for Brazilian Portuguese with tweets annotated as either toxic or non-toxic or in different types of toxicity. We present our dataset collection and annotation process, where we aimed to select candidates covering multiple demographic groups. State-of-the-art BERT models were able to achieve 76% macro-F1 score using monolingual data in the binary case. We also show that large-scale monolingual data is still needed to create more accurate models, despite recent advances in multilingual approaches. An error analysis and experiments with multi-label classification show the difficulty of classifying certain types of toxic comments that appear less frequently in our data and highlights the need to develop models that are aware of different categories of toxicity.

Carolina Scarton, Diego F. Silva, Kalina Bontcheva

Stance classification can be a powerful tool for understanding whether and which users believe in online rumours. The task aims to automatically predict the stance of replies towards a given rumour, namely support, deny, question, or comment. Numerous methods have been proposed and their performance compared in the RumourEval shared tasks in 2017 and 2019. Results demonstrated that this is a challenging problem since naturally occurring rumour stance data is highly imbalanced. This paper specifically questions the evaluation metrics used in these shared tasks. We re-evaluate the systems submitted to the two RumourEval tasks and show that the two widely adopted metrics -- accuracy and macro-F1 -- are not robust for the four-class imbalanced task of rumour stance classification, as they wrongly favour systems with highly skewed accuracy towards the majority class. To overcome this problem, we propose new evaluation metrics for rumour stance detection. These are not only robust to imbalanced data but also score higher systems that are capable of recognising the two most informative minority classes (support and deny).

Fernando Alva-Manchego, Louis Martin, Antoine Bordes et al.

In order to simplify a sentence, human editors perform multiple rewriting transformations: they split it into several shorter sentences, paraphrase words (i.e. replacing complex words or phrases by simpler synonyms), reorder components, and/or delete information deemed unnecessary. Despite these varied range of possible text alterations, current models for automatic sentence simplification are evaluated using datasets that are focused on a single transformation, such as lexical paraphrasing or splitting. This makes it impossible to understand the ability of simplification models in more realistic settings. To alleviate this limitation, this paper introduces ASSET, a new dataset for assessing sentence simplification in English. ASSET is a crowdsourced multi-reference corpus where each simplification was produced by executing several rewriting transformations. Through quantitative and qualitative experiments, we show that simplifications in ASSET are better at capturing characteristics of simplicity when compared to other standard evaluation datasets for the task. Furthermore, we motivate the need for developing better methods for automatic evaluation using ASSET, since we show that current popular metrics may not be suitable when multiple simplification transformations are performed.

Carolina Scarton, Mikel L. Forcada, Miquel Esplà-Gomis et al.

Devising metrics to assess translation quality has always been at the core of machine translation (MT) research. Traditional automatic reference-based metrics, such as BLEU, have shown correlations with human judgements of adequacy and fluency and have been paramount for the advancement of MT system development. Crowd-sourcing has popularised and enabled the scalability of metrics based on human judgements, such as subjective direct assessments (DA) of adequacy, that are believed to be more reliable than reference-based automatic metrics. Finally, task-based measurements, such as post-editing time, are expected to provide a more detailed evaluation of the usefulness of translations for a specific task. Therefore, while DA averages adequacy judgements to obtain an appraisal of (perceived) quality independently of the task, and reference-based automatic metrics try to objectively estimate quality also in a task-independent way, task-based metrics are measurements obtained either during or after performing a specific task. In this paper we argue that, although expensive, task-based measurements are the most reliable when estimating MT quality in a specific task; in our case, this task is post-editing. To that end, we report experiments on a dataset with newly-collected post-editing indicators and show their usefulness when estimating post-editing effort. Our results show that task-based metrics comparing machine-translated and post-edited versions are the best at tracking post-editing effort, as expected. These metrics are followed by DA, and then by metrics comparing the machine-translated version and independent references. We suggest that MT practitioners should be aware of these differences and acknowledge their implications when deciding how to evaluate MT for post-editing purposes.

Fernando Alva-Manchego, Louis Martin, Carolina Scarton et al.

We introduce EASSE, a Python package aiming to facilitate and standardise automatic evaluation and comparison of Sentence Simplification (SS) systems. EASSE provides a single access point to a broad range of evaluation resources: standard automatic metrics for assessing SS outputs (e.g. SARI), word-level accuracy scores for certain simplification transformations, reference-independent quality estimation features (e.g. compression ratio), and standard test data for SS evaluation (e.g. TurkCorpus). Finally, EASSE generates easy-to-visualise reports on the various metrics and features above and on how a particular SS output fares against reference simplifications. Through experiments, we show that these functionalities allow for better comparison and understanding of the performance of SS systems.

Mikel L. Forcada, Carolina Scarton, Lucia Specia et al.

A popular application of machine translation (MT) is gisting: MT is consumed as is to make sense of text in a foreign language. Evaluation of the usefulness of MT for gisting is surprisingly uncommon. The classical method uses reading comprehension questionnaires (RCQ), in which informants are asked to answer professionally-written questions in their language about a foreign text that has been machine-translated into their language. Recently, gap-filling (GF), a form of cloze testing, has been proposed as a cheaper alternative to RCQ. In GF, certain words are removed from reference translations and readers are asked to fill the gaps left using the machine-translated text as a hint. This paper reports, for thefirst time, a comparative evaluation, using both RCQ and GF, of translations from multiple MT systems for the same foreign texts, and a systematic study on the effect of variables such as gap density, gap-selection strategies, and document context in GF. The main findings of the study are: (a) both RCQ and GF clearly identify MT to be useful, (b) global RCQ and GF rankings for the MT systems are mostly in agreement, (c) GF scores vary very widely across informants, making comparisons among MT systems hard, and (d) unlike RCQ, which is framed around documents, GF evaluation can be framed at the sentence level. These findings support the use of GF as a cheaper alternative to RCQ.