Nafise Sadat Moosavi

Maggie Mi, Golzar Atefi, Atsuki Yamaguchi et al.

Idioms can be analysed in terms of their decomposability, the extent to which constituent meanings contribute to the figurative whole. Decomposability is thought to predict syntactic flexibility. Usage-based accounts instead attribute idiom behaviour to distributional experience, such as speaker familiarity and predictability. We examine these views using contextualised language models as controlled distributional learners. We propose a model-internal measure of decomposability and relate it to human ratings, syntactic flexibility, and predictability while tracking idiom learning during pretraining. Model-derived decomposability correlates weakly with human judgments and shows a small but consistent negative relationship with syntactic flexibility. Pretraining analyses show that stabilisation of idiom representations in models is not explained by frequency alone. Instead, surprisal, decomposability, and frequency all contribute, with decomposability showing the strongest training-dependent effect.

Prasetya Ajie Utama, Joshua Bambrick, Nafise Sadat Moosavi et al.

Neural abstractive summarization models are prone to generate summaries which are factually inconsistent with their source documents. Previous work has introduced the task of recognizing such factual inconsistency as a downstream application of natural language inference (NLI). However, state-of-the-art NLI models perform poorly in this context due to their inability to generalize to the target task. In this work, we show that NLI models can be effective for this task when the training data is augmented with high-quality task-oriented examples. We introduce Falsesum, a data generation pipeline leveraging a controllable text generation model to perturb human-annotated summaries, introducing varying types of factual inconsistencies. Unlike previously introduced document-level NLI datasets, our generated dataset contains examples that are diverse and inconsistent yet plausible. We show that models trained on a Falsesum-augmented NLI dataset improve the state-of-the-art performance across four benchmarks for detecting factual inconsistency in summarization. The code to obtain the dataset is available online at https://github.com/joshbambrick/Falsesum

Nafise Sadat Moosavi, Quentin Delfosse, Kristian Kersting et al.

State-of-the-art pretrained NLP models contain a hundred million to trillion parameters. Adapters provide a parameter-efficient alternative for the full finetuning in which we can only finetune lightweight neural network layers on top of pretrained weights. Adapter layers are initialized randomly. However, existing work uses the same adapter architecture -- i.e., the same adapter layer on top of each layer of the pretrained model -- for every dataset, regardless of the properties of the dataset or the amount of available training data. In this work, we introduce adaptable adapters that contain (1) learning different activation functions for different layers and different input data, and (2) a learnable switch to select and only use the beneficial adapter layers. We show that adaptable adapters achieve on-par performances with the standard adapter architecture while using a considerably smaller number of adapter layers. In addition, we show that the selected adapter architecture by adaptable adapters transfers well across different data settings and similar tasks. We propose to use adaptable adapters for designing efficient and effective adapter architectures. The resulting adapters (a) contain about 50% of the learning parameters of the standard adapter and are therefore more efficient at training and inference, and require less storage space, and (b) achieve considerably higher performances in low-data settings.

Haishuo Fang, Ji-Ung Lee, Nafise Sadat Moosavi et al.

Activation functions can have a significant impact on reducing the topological complexity of input data and therefore improve the performance of the model. Selecting a suitable activation function is an essential step in neural model design. However, the choice of activation function is seldom discussed or explored in Transformer-based language models. Their activation functions are chosen beforehand and then remain fixed from pre-training to fine-tuning. As a result, the inductive biases they imposed on models cannot be adjusted during this long life cycle. Moreover, subsequently developed models (e.g., RoBERTa, BART, and GPT-3) often follow up prior work (e.g., BERT) to use the same activation function without justification. In this paper, we investigate the effectiveness of using Rational Activation Function (RAF), a learnable activation function, in the Transformer architecture. In contrast to conventional, predefined activation functions, RAFs can adaptively learn optimal activation functions during training according to input data. Our experiments show the RAF-based Transformer (RAFT) achieves a lower validation perplexity than a vanilla BERT with the GELU function. We further evaluate RAFT on downstream tasks in low- and full-data settings. Our results show that RAFT outperforms the counterpart model across the majority of tasks and settings. For instance, RAFT outperforms vanilla BERT on the GLUE benchmark by 5.71 points on average in low-data scenario (where 100 training examples are available) and by 2.05 points on SQuAD in full-data setting. Analysis of the shapes of learned RAFs further unveils that they substantially vary between different layers of the pre-trained model and mostly look very different from conventional activation functions. RAFT opens a new research direction for analyzing and interpreting pre-trained models according to the learned activation functions.

Jan-Christoph Klie, Ji-Ung Lee, Kevin Stowe et al.

Many Natural Language Processing (NLP) systems use annotated corpora for training and evaluation. However, labeled data is often costly to obtain and scaling annotation projects is difficult, which is why annotation tasks are often outsourced to paid crowdworkers. Citizen Science is an alternative to crowdsourcing that is relatively unexplored in the context of NLP. To investigate whether and how well Citizen Science can be applied in this setting, we conduct an exploratory study into engaging different groups of volunteers in Citizen Science for NLP by re-annotating parts of a pre-existing crowdsourced dataset. Our results show that this can yield high-quality annotations and attract motivated volunteers, but also requires considering factors such as scalability, participation over time, and legal and ethical issues. We summarize lessons learned in the form of guidelines and provide our code and data to aid future work on Citizen Science.

Dominic Petrak, Nafise Sadat Moosavi, Iryna Gurevych

State-of-the-art pretrained language models tend to perform below their capabilities when applied out-of-the-box on tasks that require understanding and working with numbers. Recent work suggests two main reasons for this: (1) popular tokenisation algorithms have limited expressiveness for numbers, and (2) common pretraining objectives do not target numeracy. Approaches that address these shortcomings usually require architectural changes or pretraining from scratch. In this paper, we propose a new extended pretraining approach called Arithmetic-Based Pretraining that jointly addresses both in one extended pretraining step without requiring architectural changes or pretraining from scratch. Arithmetic-Based Pretraining combines contrastive learning to improve the number representation, and a novel extended pretraining objective called Inferable Number Prediction Task to improve numeracy. Our experiments show the effectiveness of Arithmetic-Based Pretraining in three different tasks that require improved numeracy, i.e., reading comprehension in the DROP dataset, inference-on-tables in the InfoTabs dataset, and table-to-text generation in the WikiBio and SciGen datasets.

Favour Yahdii Aghaebe, Tanefa Apekey, Elizabeth Williams et al.

Clinical interventions often hinge on age: medications and procedures safe for adults may be harmful to children or ineffective for older adults. However, as language models are increasingly integrated into biomedical evidence synthesis workflows, it remains uncertain whether these systems preserve such crucial demographic distinctions. To address this gap, we evaluate how well state-of-the-art language models retain age-related information when generating abstractive summaries of biomedical studies. We construct DemogSummary, a novel age-stratified dataset of systematic review primary studies, covering child, adult, and older adult populations. We evaluate three prominent summarisation-capable LLMs, Qwen (open-source), Longformer (open-source) and GPT-4.1 Nano (proprietary), using both standard metrics and a newly proposed Demographic Salience Score (DSS), which quantifies age-related entity retention and hallucination. Our results reveal systematic disparities across models and age groups: demographic fidelity is lowest for adult-focused summaries, and under-represented populations are more prone to hallucinations. These findings highlight the limitations of current LLMs in faithful and bias-free summarisation and point to the need for fairness-aware evaluation frameworks and summarisation pipelines in biomedical NLP.

Yiqi Liu, Nafise Sadat Moosavi, Chenghua Lin

Automatic evaluation of generated textual content presents an ongoing challenge within the field of NLP. Given the impressive capabilities of modern language models (LMs) across diverse NLP tasks, there is a growing trend to employ these models in creating innovative evaluation metrics for automated assessment of generation tasks. This paper investigates a pivotal question: Do language model-driven evaluation metrics inherently exhibit bias favoring texts generated by the same underlying language model? Specifically, we assess whether prominent LM-based evaluation metrics (e.g. BARTScore, T5Score, and GPTScore) demonstrate a favorable bias toward their respective underlying LMs in the context of summarization tasks. Our findings unveil a latent bias, particularly pronounced when such evaluation metrics are used in a reference-free manner without leveraging gold summaries. These results underscore that assessments provided by generative evaluation models can be influenced by factors beyond the inherent text quality, highlighting the necessity of developing more reliable evaluation protocols in the future.

Doan Nam Long Vu, Nafise Sadat Moosavi, Steffen Eger

The evaluation of recent embedding-based evaluation metrics for text generation is primarily based on measuring their correlation with human evaluations on standard benchmarks. However, these benchmarks are mostly from similar domains to those used for pretraining word embeddings. This raises concerns about the (lack of) generalization of embedding-based metrics to new and noisy domains that contain a different vocabulary than the pretraining data. In this paper, we examine the robustness of BERTScore, one of the most popular embedding-based metrics for text generation. We show that (a) an embedding-based metric that has the highest correlation with human evaluations on a standard benchmark can have the lowest correlation if the amount of input noise or unknown tokens increases, (b) taking embeddings from the first layer of pretrained models improves the robustness of all metrics, and (c) the highest robustness is achieved when using character-level embeddings, instead of token-based embeddings, from the first layer of the pretrained model.

Dominic Petrak, Nafise Sadat Moosavi, Ye Tian et al.

Learning from free-text human feedback is essential for dialog systems, but annotated data is scarce and usually covers only a small fraction of error types known in conversational AI. Instead of collecting and annotating new datasets from scratch, recent advances in synthetic dialog generation could be used to augment existing dialog datasets with the necessary annotations. However, to assess the feasibility of such an effort, it is important to know the types and frequency of free-text human feedback included in these datasets. In this work, we investigate this question for a variety of commonly used dialog datasets, including MultiWoZ, SGD, BABI, PersonaChat, Wizards-of-Wikipedia, and the human-bot split of the Self-Feeding Chatbot. Using our observations, we derive new taxonomies for the annotation of free-text human feedback in dialogs and investigate the impact of including such data in response generation for three SOTA language generation models, including GPT-2, LLAMA, and Flan-T5. Our findings provide new insights into the composition of the datasets examined, including error types, user response types, and the relations between them.

Silviu Paun, Juntao Yu, Nafise Sadat Moosavi et al.

Anaphoric reference is an aspect of language interpretation covering a variety of types of interpretation beyond the simple case of identity reference to entities introduced via nominal expressions covered by the traditional coreference task in its most recent incarnation in ONTONOTES and similar datasets. One of these cases that go beyond simple coreference is anaphoric reference to entities that must be added to the discourse model via accommodation, and in particular split-antecedent references to entities constructed out of other entities, as in split-antecedent plurals and in some cases of discourse deixis. Although this type of anaphoric reference is now annotated in many datasets, systems interpreting such references cannot be evaluated using the Reference coreference scorer Pradhan et al. (2014). As part of the work towards a new scorer for anaphoric reference able to evaluate all aspects of anaphoric interpretation in the coverage of the Universal Anaphora initiative, we propose in this paper a solution to the technical problem of generalizing existing metrics for identity anaphora so that they can also be used to score cases of split-antecedents. This is the first such proposal in the literature on anaphora or coreference, and has been successfully used to score both split-antecedent plural references and discourse deixis in the recent CODI/CRAC anaphora resolution in dialogue shared tasks.

Jasivan Alex Sivakumar, Nafise Sadat Moosavi

Within numerical reasoning, understanding numbers themselves is still a challenge for existing language models. Simple generalisations, such as solving 100+200 instead of 1+2, can substantially affect model performance (Sivakumar and Moosavi, 2023). Among various techniques, character-level embeddings of numbers have emerged as a promising approach to improve number representation. However, this method has limitations as it leaves the task of aggregating digit representations to the model, which lacks direct supervision for this process. In this paper, we explore the use of mathematical priors to compute aggregated digit embeddings and explicitly incorporate these aggregates into transformer models. This can be achieved either by adding a special token to the input embeddings or by introducing an additional loss function to enhance correct predictions. We evaluate the effectiveness of incorporating this explicit aggregation, analysing its strengths and shortcomings, and discuss future directions to better benefit from this approach. Our methods, while simple, are compatible with any pretrained model, easy to implement, and have been made publicly available.

Joe Stacey, Hadas Orgad, Kentaro Inui et al.

Recent work has shown that the hidden states of large language models contain signals useful for uncertainty estimation and hallucination detection, motivating a growing interest in efficient probe-based approaches. Yet it remains unclear how robust existing methods are, and which probe designs provide uncertainty estimates that are reliable under distribution shift. We present a systematic study of supervised uncertainty probes across models, tasks, and OOD settings, training over 2,000 probes while varying the representation layer, feature type, and token aggregation strategy. Our evaluation highlights poor robustness in current methods, particularly in the case of long-form generations. We also find that probe robustness is driven less by architecture and more by the probe inputs. Middle-layer representations generalise more reliably than final-layer hidden states, and aggregating across response tokens is consistently more robust than relying on single-token features. These differences are often largely invisible in-distribution but become more important under distribution shift. Informed by our evaluation, we explore a simple hybrid back-off strategy for improving robustness, arguing that better evaluation is a prerequisite for building more robust probes.

Vynska Amalia Permadi, Xingwei Tan, Nafise Sadat Moosavi et al.

Understanding culture requires reasoning across context, tradition, and implicit social knowledge, far beyond recalling isolated facts. Yet most culturally focused question answering (QA) benchmarks rely on single-hop questions, which may allow models to exploit shallow cues rather than demonstrate genuine cultural reasoning. In this work, we introduce ID-MoCQA, the first large-scale multi-hop QA dataset for assessing the cultural understanding of large language models (LLMs), grounded in Indonesian traditions and available in both English and Indonesian. We present a new framework that systematically transforms single-hop cultural questions into multi-hop reasoning chains spanning six clue types (e.g., commonsense, temporal, geographical). Our multi-stage validation pipeline, combining expert review and LLM-as-a-judge filtering, ensures high-quality question-answer pairs. Our evaluation across state-of-the-art models reveals substantial gaps in cultural reasoning, particularly in tasks requiring nuanced inference. ID-MoCQA provides a challenging and essential benchmark for advancing the cultural competency of LLMs.

Favour Yahdii Aghaebe, Tanefa Apekey, Elizabeth Williams et al.

Opinion and multi-document summarisation often involve genuinely conflicting viewpoints, yet many existing approaches, particularly LLM-based systems, implicitly smooth disagreement and over-represent majority opinions. This limits the faithfulness of generated summaries in opinion-heavy settings. We introduce a disagreement-aware synthesis pipeline that separates belief-level aggregation from language generation. Documents are first represented as structured belief sets and aggregated using distance-based belief merging operators that explicitly model conflict. Large language models are then used only to realise the aggregated beliefs as natural language summaries. We evaluate the approach across multiple model families and scales, comparing it to methods that perform explicit aggregation during generation. Our results show that while sufficiently large models can match belief-level aggregation when aggregation is handled at generation time, this behaviour is not stable across architectures or capacities. In contrast, belief-level aggregation combined with simple prompting yields consistently strong disagreement-aware performance across models, while maintaining fluent and grounded summaries.

Maggie Mi, Aline Villavicencio, Nafise Sadat Moosavi

Human processing of idioms relies on understanding the contextual sentences in which idioms occur, as well as language-intrinsic features such as frequency and speaker-intrinsic factors like familiarity. While LLMs have shown high performance on idiomaticity detection tasks, this success may be attributed to reasoning shortcuts in existing datasets. To this end, we construct a novel, controlled contrastive dataset designed to test whether LLMs can effectively use context to disambiguate idiomatic meaning. Additionally, we explore how collocational frequency and sentence probability influence model performance. Our findings reveal that LLMs often fail to resolve idiomaticity when it is required to attend to the surrounding context, and that models perform better on sentences that have higher likelihood. The collocational frequency of expressions also impacts performance. We make our code and dataset publicly available.

Kaustubh D. Dhole, Varun Gangal, Sebastian Gehrmann et al.

Data augmentation is an important component in the robustness evaluation of models in natural language processing (NLP) and in enhancing the diversity of the data they are trained on. In this paper, we present NL-Augmenter, a new participatory Python-based natural language augmentation framework which supports the creation of both transformations (modifications to the data) and filters (data splits according to specific features). We describe the framework and an initial set of 117 transformations and 23 filters for a variety of natural language tasks. We demonstrate the efficacy of NL-Augmenter by using several of its transformations to analyze the robustness of popular natural language models. The infrastructure, datacards and robustness analysis results are available publicly on the NL-Augmenter repository (https://github.com/GEM-benchmark/NL-Augmenter).

Nafise Sadat Moosavi, Andreas Rücklé, Dan Roth et al.

In this paper, we introduce SciGen, a new challenge dataset for the task of reasoning-aware data-to-text generation consisting of tables from scientific articles and their corresponding descriptions. Describing scientific tables goes beyond the surface realization of the table content and requires reasoning over table values. The unique properties of SciGen are that (1) tables mostly contain numerical values, and (2) the corresponding descriptions require arithmetic reasoning. SciGen is therefore the first dataset that assesses the arithmetic reasoning capabilities of generation models on complex input structures, i.e., tables from scientific articles. We study the effectiveness of state-of-the-art data-to-text generation models on SciGen and evaluate the results using common metrics as well as human evaluation. Our results and analyses show that (a) while humans like to reason for describing scientific tables, the ability of state-of-the-art models is severely limited on this task, (b) while adding more training data improves the results, it is not the solution for reasoning-aware text generation, and (c) one of the main bottlenecks for this task is the lack of proper automatic evaluation metrics. The data, code, and annotations for human evaluation will be available at https://github.com/UKPLab/SciGen. SciGen opens new avenues for future research in reasoning-aware text generation and evaluation.

Mingzhu Wu, Nafise Sadat Moosavi, Dan Roth et al.

Coreference resolution is essential for natural language understanding and has been long studied in NLP. In recent years, as the format of Question Answering (QA) became a standard for machine reading comprehension (MRC), there have been data collection efforts, e.g., Dasigi et al. (2019), that attempt to evaluate the ability of MRC models to reason about coreference. However, as we show, coreference reasoning in MRC is a greater challenge than earlier thought; MRC datasets do not reflect the natural distribution and, consequently, the challenges of coreference reasoning. Specifically, success on these datasets does not reflect a model's proficiency in coreference reasoning. We propose a methodology for creating MRC datasets that better reflect the challenges of coreference reasoning and use it to create a sample evaluation set. The results on our dataset show that state-of-the-art models still struggle with these phenomena. Furthermore, we develop an effective way to use naturally occurring coreference phenomena from existing coreference resolution datasets when training MRC models. This allows us to show an improvement in the coreference reasoning abilities of state-of-the-art models. The code and the resulting dataset are available at https://github.com/UKPLab/coref-reasoning-in-qa.

Nafise Sadat Moosavi, Marcel de Boer, Prasetya Ajie Utama et al.

Existing NLP datasets contain various biases, and models tend to quickly learn those biases, which in turn limits their robustness. Existing approaches to improve robustness against dataset biases mostly focus on changing the training objective so that models learn less from biased examples. Besides, they mostly focus on addressing a specific bias, and while they improve the performance on adversarial evaluation sets of the targeted bias, they may bias the model in other ways, and therefore, hurt the overall robustness. In this paper, we propose to augment the input sentences in the training data with their corresponding predicate-argument structures, which provide a higher-level abstraction over different realizations of the same meaning and help the model to recognize important parts of sentences. We show that without targeting a specific bias, our sentence augmentation improves the robustness of transformer models against multiple biases. In addition, we show that models can still be vulnerable to the lexical overlap bias, even when the training data does not contain this bias, and that the sentence augmentation also improves the robustness in this scenario. We will release our adversarial datasets to evaluate bias in such a scenario as well as our augmentation scripts at https://github.com/UKPLab/data-augmentation-for-robustness.

Nafise Sadat Moosavi, Leo Born, Massimo Poesio et al.

The common practice in coreference resolution is to identify and evaluate the maximum span of mentions. The use of maximum spans tangles coreference evaluation with the challenges of mention boundary detection like prepositional phrase attachment. To address this problem, minimum spans are manually annotated in smaller corpora. However, this additional annotation is costly and therefore, this solution does not scale to large corpora. In this paper, we propose the MINA algorithm for automatically extracting minimum spans to benefit from minimum span evaluation in all corpora. We show that the extracted minimum spans by MINA are consistent with those that are manually annotated by experts. Our experiments show that using minimum spans is in particular important in cross-dataset coreference evaluation, in which detected mention boundaries are noisier due to domain shift. We will integrate MINA into https://github.com/ns-moosavi/coval for reporting standard coreference scores based on both maximum and automatically detected minimum spans.

Steffen Eger, Yong Cao, Jennifer D'Souza et al.

With the advent of large multimodal language models, science is now at a threshold of an AI-based technological transformation. Recently, a plethora of new AI models and tools has been proposed, promising to empower researchers and academics worldwide to conduct their research more effectively and efficiently. This includes all aspects of the research cycle, especially (1) searching for relevant literature; (2) generating research ideas and conducting experimentation; generating (3) text-based and (4) multimodal content (e.g., scientific figures and diagrams); and (5) AI-based automatic peer review. In this survey, we provide an in-depth overview over these exciting recent developments, which promise to fundamentally alter the scientific research process for good. Our survey covers the five aspects outlined above, indicating relevant datasets, methods and results (including evaluation) as well as limitations and scope for future research. Ethical concerns regarding shortcomings of these tools and potential for misuse (fake science, plagiarism, harms to research integrity) take a particularly prominent place in our discussion. We hope that our survey will not only become a reference guide for newcomers to the field but also a catalyst for new AI-based initiatives in the area of "AI4Science".

Valeria Pastorino, Jasivan A. Sivakumar, Nafise Sadat Moosavi

Previous studies on framing have relied on manual analysis or fine-tuning models with limited annotated datasets. However, pre-trained models, with their diverse training backgrounds, offer a promising alternative. This paper presents a comprehensive analysis of GPT-4, GPT-3.5 Turbo, and FLAN-T5 models in detecting framing in news headlines. We evaluated these models in various scenarios: zero-shot, few-shot with in-domain examples, cross-domain examples, and settings where models explain their predictions. Our results show that explainable predictions lead to more reliable outcomes. GPT-4 performed exceptionally well in few-shot settings but often misinterpreted emotional language as framing, highlighting a significant challenge. Additionally, the results suggest that consistent predictions across multiple models could help identify potential annotation inaccuracies in datasets. Finally, we propose a new small dataset for real-world evaluation on headlines from a diverse set of topics.

Hamidreza Saffari, Mohammadamin Shafiei, Hezhao Zhang et al.

Dehumanization, i.e., denying human qualities to individuals or groups, is a particularly harmful form of hate speech that can normalize violence against marginalized communities. Despite advances in NLP for detecting general hate speech, approaches to identifying dehumanizing language remain limited due to scarce annotated data and the subtle nature of such expressions. In this work, we systematically evaluate four state-of-the-art large language models (LLMs) - Claude, GPT, Mistral, and Qwen - for dehumanization detection. Our results show that only one model-Claude-achieves strong performance (over 80% F1) under an optimized configuration, while others, despite their capabilities, perform only moderately. Performance drops further when distinguishing dehumanization from related hate types such as derogation. We also identify systematic disparities across target groups: models tend to over-predict dehumanization for some identities (e.g., Gay men), while under-identifying it for others (e.g., Refugees). These findings motivate the need for systematic, group-level evaluation when applying pretrained language models to dehumanization detection tasks.

Xiao Wang, Daniil Larionov, Siwei Wu et al.

Evaluating the quality of generated text automatically remains a significant challenge. Conventional reference-based metrics have been shown to exhibit relatively weak correlation with human evaluations. Recent research advocates the use of large language models (LLMs) as source-based metrics for natural language generation (NLG) assessment. While promising, LLM-based metrics, particularly those using smaller models, still fall short in aligning with human judgments. In this work, we introduce ContrastScore, a contrastive evaluation metric designed to enable higher-quality, less biased, and more efficient assessment of generated text. We evaluate ContrastScore on two NLG tasks: machine translation and summarization. Experimental results show that ContrastScore consistently achieves stronger correlation with human judgments than both single-model and ensemble-based baselines. Notably, ContrastScore based on Qwen 3B and 0.5B even outperforms Qwen 7B, despite having only half as many parameters, demonstrating its efficiency. Furthermore, it effectively mitigates common evaluation biases such as length and likelihood preferences, resulting in more robust automatic evaluation.

Ian W. Kennedy, Nafise Sadat Moosavi

Additive quantization enables extreme LLM compression with O(1) lookup-table dequantization, making it attractive for edge deployment. Yet at 2-bit precision, it often fails catastrophically, even with extensive search and finetuning. We show that the dominant bottleneck is codebook initialisation. Greedy sequential initialisation frequently places the model in poor optimisation regions that subsequent beam search and PV-tuning struggle to overcome. We analyse this behaviour through the representational ratio \r{ho} = N/KM, which characterises the relationship between weight groups and codebook capacity, and propose OA-EM, an output-aware EM initialisation method using Hessian-weighted Mahalanobis distance. Across compression rates, search budgets, and three architectures (Llama 3.2 3B, Llama 3.1 8B, Qwen 2.5 3B), OA-EM consistently produces better solutions after PV-tuning and dominates the quality-compute frontier. The severity of the bottleneck scales with \r{ho}: moderate at 3 bpp but extreme at 2 bpp, where poor initialisation can degrade perplexity by orders of magnitude. More broadly, our results highlight the importance of optimisation geometry in compressed model spaces, where initialisation can dominate subsequent search and fine-tuning.

Hend ElGhazaly, Bahman Mirheidari, Nafise Sadat Moosavi et al.

This study investigates factors influencing Automatic Speech Recognition (ASR) systems' fairness and performance across genders, beyond the conventional examination of demographics. Using the LibriSpeech dataset and the Whisper small model, we analyze how performance varies across different gender representations in training data. Our findings suggest a complex interplay between the gender ratio in training data and ASR performance. Optimal fairness occurs at specific gender distributions rather than a simple 50-50 split. Furthermore, our findings suggest that factors like pitch variability can significantly affect ASR accuracy. This research contributes to a deeper understanding of biases in ASR systems, highlighting the importance of carefully curated training data in mitigating gender bias.

Mugdha Pandya, Nafise Sadat Moosavi, Diana Maynard

Resolving disagreement in manual annotation typically consists of removing unreliable annotators and using a label aggregation strategy such as majority vote or expert opinion to resolve disagreement. These may have the side-effect of silencing or under-representing minority but equally valid opinions. In this paper, we study the impact of standard label aggregation strategies on minority opinion representation in sexism detection. We investigate the quality and value of minority annotations, and then examine their effect on the class distributions in gold labels, as well as how this affects the behaviour of models trained on the resulting datasets. Finally, we discuss the potential biases introduced by each method and how they can be amplified by the models.

Joseph James, Chenghao Xiao, Yucheng Li et al.

Scientific rigour tends to be sidelined in favour of bold statements, leading authors to overstate claims beyond what their results support. We present RIGOURATE, a two-stage multimodal framework that retrieves supporting evidence from a paper's body and assigns each claim an overstatement score. The framework consists of a dataset of over 10K claim-evidence sets from ICLR and NeurIPS papers, annotated using eight LLMs, with overstatement scores calibrated using peer-review comments and validated through human evaluation. It employes a fine-tuned reranker for evidence retrieval and a fine-tuned model to predict overstatement scores with justification. Compared to strong baselines, RIGOURATE enables improved evidence retrieval and overstatement detection. Overall, our work operationalises evidential proportionality and supports clearer, more transparent scientific communication.

Huiyin Xue, Nafise Sadat Moosavi, Nikolaos Aletras

The success of Transformer language models is widely credited to their dot-product attention mechanism, which interweaves a set of key design principles: mixing information across positions (enabling multi-token interactions), sequence-dependent activations (where attention weights adapt to each input), a specific mathematical form (dot-product similarities plus softmax weighting), and coupling of queries and keys to evolving hidden states (grounding attention in the current layer). However, the necessity of each of these principles remains largely untested. In this work, we systematically deconstruct attention by designing controlled variants that selectively relax these principles, applied both uniformly across all layers and in hybrid architectures where only some layers retain standard attention. Our empirical analysis reveals that mechanisms for mixing tokens are indispensable, as their absence collapses models to near-random behavior, while the exact mathematical form and sequence dependency can be substantially relaxed, especially when preserved in just a subset of layers. Surprisingly, even variants that fail in isolation can achieve robust performance when interleaved with standard attention, highlighting a cooperative effect. These findings deepen our understanding of what truly underpins attention's effectiveness and open new avenues for simplifying language models without sacrificing performance.

Maggie Mi, Aline Villavicencio, Nafise Sadat Moosavi

Language models often struggle with idiomatic, figurative, or context-sensitive inputs, not because they produce flawed outputs, but because they misinterpret the input from the outset. We propose an input-only method for anticipating such failures using token-level likelihood features inspired by surprisal and the Uniform Information Density hypothesis. These features capture localized uncertainty in input comprehension and outperform standard baselines across five linguistically challenging datasets. We show that span-localized features improve error detection for larger models, while smaller models benefit from global patterns. Our method requires no access to outputs or hidden activations, offering a lightweight and generalizable approach to pre-generation error prediction.

Mohammadamin Shafiei, Hamidreza Saffari, Nafise Sadat Moosavi

Large language models (LLMs) are known to be sensitive to input phrasing, but the mechanisms by which semantic cues shape reasoning remain poorly understood. We investigate this phenomenon in the context of comparative math problems with objective ground truth, revealing a consistent and directional framing bias: logically equivalent questions containing the words ``more'', ``less'', or ``equal'' systematically steer predictions in the direction of the framing term. To study this effect, we introduce MathComp, a controlled benchmark of 300 comparison scenarios, each evaluated under 14 prompt variants across three LLM families. We find that model errors frequently reflect linguistic steering, systematic shifts toward the comparative term present in the prompt. Chain-of-thought prompting reduces these biases, but its effectiveness varies: free-form reasoning is more robust, while structured formats may preserve or reintroduce directional drift. Finally, we show that including demographic identity terms (e.g., ``a woman'', ``a Black person'') in input scenarios amplifies directional drift, despite identical underlying quantities, highlighting the interplay between semantic framing and social referents. These findings expose critical blind spots in standard evaluation and motivate framing-aware benchmarks for diagnosing reasoning robustness and fairness in LLMs.

Mohammadamin Shafiei, Hamidreza Saffari, Nafise Sadat Moosavi

As Large Language Models are increasingly deployed in high-stakes domains, their ability to detect false assumptions and reason critically is crucial for ensuring reliable outputs. False-premise questions (FPQs) serve as an important evaluation method by exposing cases where flawed assumptions lead to incorrect responses. While existing benchmarks focus on single-hop FPQs, real-world reasoning often requires multi-hop inference, where models must verify consistency across multiple reasoning steps rather than relying on surface-level cues. To address this gap, we introduce MultiHoax, a benchmark for evaluating LLMs' ability to handle false premises in complex, multi-step reasoning tasks. Our dataset spans seven countries and ten diverse knowledge categories, using Wikipedia as the primary knowledge source to enable factual reasoning across regions. Experiments reveal that state-of-the-art LLMs struggle to detect false premises across different countries, knowledge categories, and multi-hop reasoning types, highlighting the need for improved false premise detection and more robust multi-hop reasoning capabilities in LLMs.

Valeria Pastorino, Nafise Sadat Moosavi

Framing in media critically shapes public perception by selectively emphasizing some details while downplaying others. With the rise of large language models in automated news and content creation, there is growing concern that these systems may introduce or even amplify framing biases compared to human authors. In this paper, we explore how framing manifests in both out-of-the-box and fine-tuned LLM-generated news content. Our analysis reveals that, particularly in politically and socially sensitive contexts, LLMs tend to exhibit more pronounced framing than their human counterparts. In addition, we observe significant variation in framing tendencies across different model architectures, with some models displaying notably higher biases. These findings point to the need for effective post-training mitigation strategies and tighter evaluation frameworks to ensure that automated news content upholds the standards of balanced reporting.

Jasivan Alex Sivakumar, Nafise Sadat Moosavi

While pre-trained language models achieve impressive performance on various NLP benchmarks, they still struggle with tasks that require numerical reasoning. Recent advances in improving numerical reasoning are mostly achieved using very large language models that contain billions of parameters and are not accessible to everyone. In addition, numerical reasoning is measured using a single score on existing datasets. As a result, we do not have a clear understanding of the strengths and shortcomings of existing models on different numerical reasoning aspects and therefore, potential ways to improve them apart from scaling them up. Inspired by CheckList (Ribeiro et al., 2020), we introduce a multi-view evaluation set for numerical reasoning in English, called FERMAT. Instead of reporting a single score on a whole dataset, FERMAT evaluates models on various key numerical reasoning aspects such as number understanding, mathematical operations, and training dependency. Apart from providing a comprehensive evaluation of models on different numerical reasoning aspects, FERMAT enables a systematic and automated generation of an arbitrarily large training or evaluation set for each aspect.The datasets and codes are publicly available to generate further multi-view data for ulterior tasks and languages.

Prasetya Ajie Utama, Nafise Sadat Moosavi, Victor Sanh et al.

Recent prompt-based approaches allow pretrained language models to achieve strong performances on few-shot finetuning by reformulating downstream tasks as a language modeling problem. In this work, we demonstrate that, despite its advantages on low data regimes, finetuned prompt-based models for sentence pair classification tasks still suffer from a common pitfall of adopting inference heuristics based on lexical overlap, e.g., models incorrectly assuming a sentence pair is of the same meaning because they consist of the same set of words. Interestingly, we find that this particular inference heuristic is significantly less present in the zero-shot evaluation of the prompt-based model, indicating how finetuning can be destructive to useful knowledge learned during the pretraining. We then show that adding a regularization that preserves pretraining weights is effective in mitigating this destructive tendency of few-shot finetuning. Our evaluation on three datasets demonstrates promising improvements on the three corresponding challenge datasets used to diagnose the inference heuristics.

Juntao Yu, Nafise Sadat Moosavi, Silviu Paun et al.

The state-of-the-art on basic, single-antecedent anaphora has greatly improved in recent years. Researchers have therefore started to pay more attention to more complex cases of anaphora such as split-antecedent anaphora, as in Time-Warner is considering a legal challenge to Telecommunications Inc's plan to buy half of Showtime Networks Inc-a move that could lead to all-out war between the two powerful companies. Split-antecedent anaphora is rarer and more complex to resolve than single-antecedent anaphora; as a result, it is not annotated in many datasets designed to test coreference, and previous work on resolving this type of anaphora was carried out in unrealistic conditions that assume gold mentions and/or gold split-antecedent anaphors are available. These systems also focus on split-antecedent anaphors only. In this work, we introduce a system that resolves both single and split-antecedent anaphors, and evaluate it in a more realistic setting that uses predicted mentions. We also start addressing the question of how to evaluate single and split-antecedent anaphors together using standard coreference evaluation metrics.

Juntao Yu, Nafise Sadat Moosavi, Silviu Paun et al.

Now that the performance of coreference resolvers on the simpler forms of anaphoric reference has greatly improved, more attention is devoted to more complex aspects of anaphora. One limitation of virtually all coreference resolution models is the focus on single-antecedent anaphors. Plural anaphors with multiple antecedents-so-called split-antecedent anaphors (as in John met Mary. They went to the movies) have not been widely studied, because they are not annotated in ONTONOTES and are relatively infrequent in other corpora. In this paper, we introduce the first model for unrestricted resolution of split-antecedent anaphors. We start with a strong baseline enhanced by BERT embeddings, and show that we can substantially improve its performance by addressing the sparsity issue. To do this, we experiment with auxiliary corpora where split-antecedent anaphors were annotated by the crowd, and with transfer learning models using element-of bridging references and single-antecedent coreference as auxiliary tasks. Evaluation on the gold annotated ARRAU corpus shows that the out best model uses a combination of three auxiliary corpora achieved F1 scores of 70% and 43.6% when evaluated in a lenient and strict setting, respectively, i.e., 11 and 21 percentage points gain when compared with our baseline.

Mingzhu Wu, Nafise Sadat Moosavi, Andreas Rücklé et al.

Existing NLP datasets contain various biases that models can easily exploit to achieve high performances on the corresponding evaluation sets. However, focusing on dataset-specific biases limits their ability to learn more generalizable knowledge about the task from more general data patterns. In this paper, we investigate the impact of debiasing methods for improving generalization and propose a general framework for improving the performance on both in-domain and out-of-domain datasets by concurrent modeling of multiple biases in the training data. Our framework weights each example based on the biases it contains and the strength of those biases in the training data. It then uses these weights in the training objective so that the model relies less on examples with high bias weights. We extensively evaluate our framework on extractive question answering with training data from various domains with multiple biases of different strengths. We perform the evaluations in two different settings, in which the model is trained on a single domain or multiple domains simultaneously, and show its effectiveness in both settings compared to state-of-the-art debiasing methods.

Prasetya Ajie Utama, Nafise Sadat Moosavi, Iryna Gurevych

NLU models often exploit biases to achieve high dataset-specific performance without properly learning the intended task. Recently proposed debiasing methods are shown to be effective in mitigating this tendency. However, these methods rely on a major assumption that the types of bias should be known a-priori, which limits their application to many NLU tasks and datasets. In this work, we present the first step to bridge this gap by introducing a self-debiasing framework that prevents models from mainly utilizing biases without knowing them in advance. The proposed framework is general and complementary to the existing debiasing methods. We show that it allows these existing methods to retain the improvement on the challenge datasets (i.e., sets of examples designed to expose models' reliance on biases) without specifically targeting certain biases. Furthermore, the evaluation suggests that applying the framework results in improved overall robustness.

Prasetya Ajie Utama, Nafise Sadat Moosavi, Iryna Gurevych

Models for natural language understanding (NLU) tasks often rely on the idiosyncratic biases of the dataset, which make them brittle against test cases outside the training distribution. Recently, several proposed debiasing methods are shown to be very effective in improving out-of-distribution performance. However, their improvements come at the expense of performance drop when models are evaluated on the in-distribution data, which contain examples with higher diversity. This seemingly inevitable trade-off may not tell us much about the changes in the reasoning and understanding capabilities of the resulting models on broader types of examples beyond the small subset represented in the out-of-distribution data. In this paper, we address this trade-off by introducing a novel debiasing method, called confidence regularization, which discourage models from exploiting biases while enabling them to receive enough incentive to learn from all the training examples. We evaluate our method on three NLU tasks and show that, in contrast to its predecessors, it improves the performance on out-of-distribution datasets (e.g., 7pp gain on HANS dataset) while maintaining the original in-distribution accuracy.

Andreas Rücklé, Nafise Sadat Moosavi, Iryna Gurevych

Supervised training of neural models to duplicate question detection in community Question Answering (cQA) requires large amounts of labeled question pairs, which are costly to obtain. To minimize this cost, recent works thus often used alternative methods, e.g., adversarial domain adaptation. In this work, we propose two novel methods: (1) the automatic generation of duplicate questions, and (2) weak supervision using the title and body of a question. We show that both can achieve improved performances even though they do not require any labeled data. We provide comprehensive comparisons of popular training strategies, which provides important insights on how to best train models in different scenarios. We show that our proposed approaches are more effective in many cases because they can utilize larger amounts of unlabeled data from cQA forums. Finally, we also show that our proposed approach for weak supervision with question title and body information is also an effective method to train cQA answer selection models without direct answer supervision.

Nafise Sadat Moosavi, Prasetya Ajie Utama, Andreas Rücklé et al.

The task of natural language inference (NLI) is to identify the relation between the given premise and hypothesis. While recent NLI models achieve very high performance on individual datasets, they fail to generalize across similar datasets. This indicates that they are solving NLI datasets instead of the task itself. In order to improve generalization, we propose to extend the input representations with an abstract view of the relation between the hypothesis and the premise, i.e., how well the individual words, or word n-grams, of the hypothesis are covered by the premise. Our experiments show that the use of this information considerably improves generalization across different NLI datasets without requiring any external knowledge or additional data. Finally, we show that using the coverage information is not only beneficial for improving the performance across different datasets of the same task. The resulting generalization improves the performance across datasets that belong to similar but not the same tasks.

Nafise Sadat Moosavi, Michael Strube

Coreference resolution is an intermediate step for text understanding. It is used in tasks and domains for which we do not necessarily have coreference annotated corpora. Therefore, generalization is of special importance for coreference resolution. However, while recent coreference resolvers have notable improvements on the CoNLL dataset, they struggle to generalize properly to new domains or datasets. In this paper, we investigate the role of linguistic features in building more generalizable coreference resolvers. We show that generalization improves only slightly by merely using a set of additional linguistic features. However, employing features and subsets of their values that are informative for coreference resolution, considerably improves generalization. Thanks to better generalization, our system achieves state-of-the-art results in out-of-domain evaluations, e.g., on WikiCoref, our system, which is trained on CoNLL, achieves on-par performance with a system designed for this dataset.

Benjamin Heinzerling, Nafise Sadat Moosavi, Michael Strube

Selectional preferences have long been claimed to be essential for coreference resolution. However, they are mainly modeled only implicitly by current coreference resolvers. We propose a dependency-based embedding model of selectional preferences which allows fine-grained compatibility judgments with high coverage. We show that the incorporation of our model improves coreference resolution performance on the CoNLL dataset, matching the state-of-the-art results of a more complex system. However, it comes with a cost that makes it debatable how worthwhile such improvements are.

Nafise Sadat Moosavi, Michael Strube

Lexical features are a major source of information in state-of-the-art coreference resolvers. Lexical features implicitly model some of the linguistic phenomena at a fine granularity level. They are especially useful for representing the context of mentions. In this paper we investigate a drawback of using many lexical features in state-of-the-art coreference resolvers. We show that if coreference resolvers mainly rely on lexical features, they can hardly generalize to unseen domains. Furthermore, we show that the current coreference resolution evaluation is clearly flawed by only evaluating on a specific split of a specific dataset in which there is a notable overlap between the training, development and test sets.

Nafise Sadat Moosavi, Michael Strube

Only a year ago, all state-of-the-art coreference resolvers were using an extensive amount of surface features. Recently, there was a paradigm shift towards using word embeddings and deep neural networks, where the use of surface features is very limited. In this paper, we show that a simple SVM model with surface features outperforms more complex neural models for detecting anaphoric mentions. Our analysis suggests that using generalized representations and surface features have different strength that should be both taken into account for improving coreference resolution.