Papers by Nafise Sadat Moosavi
Fairness in Automatic Speech Recognition Isn’t a One-Size-Fits-All (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Pre-trained speech models like Whisper exhibit inconsistent group-level performance that varies across domains. |
| Approach: | They fine-tune a Whisper model on the Fair-Speech corpus using basic fine- tuning, demographic rebalancing, gender-swapped data augmentation and a novel contrastive learning objective. |
| Outcome: | The proposed method achieves stable, cross-domain fairness improvements without changes to the training data distribution and with minimal accuracy trade-offs. |
Avoiding Inference Heuristics in Few-shot Prompt-based Finetuning (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent prompt-based approaches allow pretrained language models to achieve strong performances on few-shot finetuning by reformulating downstream task instances as a language modeling problem. |
| Approach: | They propose to reformulate downstream tasks as a language modeling problem and add a regularization that preserves pretraining weights to the model to mitigate the destructive tendency of few-shot finetuning. |
| Outcome: | The proposed model performs better on low data regimes than the standard model on few-shot finetuning. |
The Universal Anaphora Scorer (2022.lrec-1)
Copied to clipboard
| Challenge: | a new version of the Reference Coreference Scorer is proposed to evaluate anaphoric interpretations . the proposed approach to evaluation of split antecedent anaphorisms is entirely novel . |
| Approach: | They propose an extended version of the Reference Coreference Scorer to evaluate anaphoric interpretations . the UA scorer supports the evaluation of split antecedent anaphorisms and discourse deixis . |
| Outcome: | The proposed method can be used to evaluate anaphoric interpretations in an extended range of anas . it supports evaluations of split antecedent anaphorisms and discourse deixis, for which no tools exist . |
RIGOURATE: Quantifying Scientific Exaggeration with Evidence-Aligned Claim Evaluation (2026.findings-acl)
Copied to clipboard
| Challenge: | Scientific rigour tends to be sidelined in favour of bold statements, leading authors to overstate claims beyond what their results support. |
| Approach: | They propose a multimodal framework that retrieves supporting evidence from a paper and assigns each claim an overstatement score. |
| Outcome: | The proposed framework retrieves supporting evidence from ICLR and NeurIPS papers and assigns each claim an overstatement score. |
Coreference Reasoning in Machine Reading Comprehension (2021.acl-long)
Copied to clipboard
| Challenge: | Existing datasets for machine reading comprehension do not reflect the natural distribution and, consequently, the challenges of coreference reasoning. |
| Approach: | They propose to use existing coreference resolution datasets to train machine reading comprehension models to better reflect the natural distribution and, consequently, the challenges of coreference reasoning. |
| Outcome: | The proposed method improves the performance of state-of-the-art models on a set of coreference-related datasets. |
Towards Debiasing NLU Models from Unknown Biases (2020.emnlp-main)
Copied to clipboard
| Challenge: | Recent proposed debiasing methods rely on the assumption that the types of bias should be known a-priori, which limits their application to many NLU tasks and datasets. |
| Approach: | They propose a framework that prevents models from mainly utilizing biases without knowing them in advance. |
| Outcome: | The proposed framework allows existing methods to retain performance improvement on challenge datasets without specifically targeting biases. |
Mind the Trade-off: Debiasing NLU Models without Degrading the In-distribution Performance (2020.acl-main)
Copied to clipboard
| Challenge: | Recent studies show that pre-trained language models rely heavily on idiosyncratic biases of datasets. |
| Approach: | They propose a method which discourages models from exploiting biases while enabling them to receive enough incentive to learn from all the training examples. |
| Outcome: | The proposed method improves on out-of-distribution datasets while maintaining original in-district accuracy. |
Stay Together: A System for Single and Split-antecedent Anaphora Resolution (2021.naacl-main)
Copied to clipboard
| Challenge: | Recent work on single-antecedent anaphora has greatly improved . attention has now turned to more complex cases of anaphorisms such as split-antevore anaprs . |
| Approach: | They propose a system that resolves both single and split-antecedent anaphors and evaluates it in a more realistic setting that uses predicted mentions. |
| Outcome: | The proposed system resolves both single and split-antecedent anaphora and evaluates it in a more realistic setting that uses predicted mentions. |
Neural Duplicate Question Detection without Labeled Training Data (D19-1)
Copied to clipboard
| Challenge: | Recent studies have used alternative methods to train neural models to duplicate question detection in community Question Answering forums. |
| Approach: | They propose two new methods for supervised question detection in community Question Answering forums . they propose weak supervision using title and body of question and automatic generation of duplicate questions . |
| Outcome: | The proposed methods can achieve better performance even without labeled data. |
Transformers with Learnable Activation Functions (2023.findings-eacl)
Copied to clipboard
| Challenge: | Activation functions can reduce the topological complexity of input data and improve model performance. |
| Approach: | They propose to consider data as a topology with its own shape to simplify its complexity and make it linearly separable in the output space. |
| Outcome: | The RAF-based Transformer model outperforms its FAF-based counterpart on the GLUE benchmark by 5.71 points and 2.05 points on SQuAD with all available data. |
FERMAT: An Alternative to Accuracy for Numerical Reasoning (2023.acl-long)
Copied to clipboard
| Challenge: | Existing numerical reasoning models are too weak for downstream tasks like fact-checking . FERMAT evaluates models on number understanding, mathematical operations, and training dependency . |
| Approach: | They propose a multi-view evaluation set for numerical reasoning in English that evaluates models on key numerical reasoning aspects instead of reporting a single score on a whole dataset. |
| Outcome: | FERMAT evaluates models on number understanding, mathematical operations, and training dependency. |
Rethinking the Idiomaticity Decomposability Hypothesis: Evidence from Distributional Learning (2026.acl-long)
Copied to clipboard
| Challenge: | Decomposability is thought to predict syntactic flexibility, but is not attributed to distributional experience. |
| Approach: | They propose a model-internal measure of decomposability and relate it to human ratings, syntactic flexibility, and predictability while tracking idiom learning during pretraining. |
| Outcome: | The proposed model-internal measure correlates weakly with human judgments and shows a small but consistent negative relationship with syntactic flexibility. |
MultiHoax: A Dataset of Multi-hop False-premise questions (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks focus on single-hop FPQs, but real-world reasoning often requires multi-hop inference . state-of-the-art LLMs struggle to detect false premises across different countries, knowledge categories, and multi-step reasoning types. |
| Approach: | They propose a benchmark to evaluate Large Language Models' ability to handle false premises in complex, multi-step reasoning tasks. |
| Outcome: | The proposed tests show that state-of-the-art LLMs struggle to detect false premises across different countries, knowledge categories, and multi-hop reasoning types. |
Arithmetic-Based Pretraining Improving Numeracy of Pretrained Language Models (2023.starsem-1)
Copied to clipboard
| Challenge: | Recent work suggests that pretrained language models perform below their capabilities when applied out-of-the-box on tasks that require understanding and working with numbers. |
| Approach: | They propose an extended pretraining approach that addresses both in one extended step . they propose a novel extended pre training objective called Inferable Number Prediction Task to improve numeracy. |
| Outcome: | The proposed approach improves reading comprehension and inference-on-tables tasks without architectural changes or pretraining from scratch. |
Universal Anaphora: The First Three Years (2024.lrec-main)
Copied to clipboard
Massimo Poesio, Maciej Ogrodniczuk, Vincent Ng, Sameer Pradhan, Juntao Yu, Nafise Sadat Moosavi, Silviu Paun, Amir Zeldes, Anna Nedoluzhko, Michal Novák, Martin Popel, Zdeněk Žabokrtský, Daniel Zeman
| Challenge: | Universal Anaphora initiative aims to push forward the state of the art in anaphora and anaphorism resolution by expanding the aspects of anaphonic interpretation which are or can be reliably annotated in an anagraphic corpora. |
| Approach: | They propose to develop a standard for anaphoric annotations and a method for evaluating models that can carry out this type of interpretation. |
| Outcome: | The Universal Anaphora initiative aims to push forward the state of the art in anaphora and anaphorism resolution by producing unified standards to annotate and encode annotations, delivering datasets encoded according to these standards, and developing methods for evaluating models that carry out this type of interpretation. |
Free the Plural: Unrestricted Split-Antecedent Anaphora Resolution (2020.coling-main)
Copied to clipboard
| Challenge: | a limitation of coreference resolution models is the focus on single-antecedent anaphors. |
| Approach: | They propose a model for unrestricted resolution of split-antecedent anaphors with multiple antecedents . they use auxiliary corpora where split-antcedent ananaphor was annotated by crowd . |
| Outcome: | The proposed model significantly improves on a baseline enhanced by BERT embeddings on anaphoric reference corpus. |
How to Leverage Digit Embeddings to Represent Numbers? (2025.coling-main)
Copied to clipboard
| Challenge: | Existing numerical reasoning models struggle to understand numbers, despite simple generalisations. |
| Approach: | They propose to use mathematical priors to compute digit embeddings and explicitly incorporate them into transformer models by adding a special token to the input embedded digits or introducing an additional loss function to enhance correct predictions. |
| Outcome: | The proposed method is compatible with any pretrained model and easy to implement. |
Improving QA Generalization by Concurrent Modeling of Multiple Biases (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing approaches for debiasing datasets are weaker than current approaches for generalization. |
| Approach: | They propose a framework for analyzing multiple biases in training data to reduce bias weighting. |
| Outcome: | The proposed framework improves generalization on in-domain and out-of-domain datasets by weighting examples based on their strengths and bias strengths. |
Using Automatically Extracted Minimum Spans to Disentangle Coreference Evaluation from Boundary Detection (P19-1)
Copied to clipboard
| Challenge: | Existing methods to evaluate maximum spans tangle coreference evaluation with mention boundary detection . however, this method is costly and does not scale to large corpora. |
| Approach: | They propose an algorithm for automatically extracting minimum spans to benefit from minimum span evaluation in all corpora. |
| Outcome: | The proposed algorithm is consistent with those manually annotated by experts. |
Beyond Hate Speech: NLP’s Challenges and Opportunities in Uncovering Dehumanizing Language (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing hate speech datasets rarely contain enough instances of dehumanizing content, and current models struggle to distinguish such language from more benign forms of hate or offense. |
| Approach: | They evaluate four state-of-the-art large language models for dehumanization detection. |
| Outcome: | The proposed models perform only moderately under an optimized configuration, while others over-predict dehumanization for some identities, while under-identifying it for others. |
Using Linguistic Features to Improve the Generalization Capability of Neural Coreference Resolvers (D18-1)
Copied to clipboard
| Challenge: | Recent coreference resolvers have notable improvements on the CoNLL evaluation sets, but struggle to generalize properly to new datasets. |
| Approach: | They investigate the role of linguistic features in building more generalizable coreference resolvers . they show that employing features and subsets of their values that are informative for coreference resolution improves generalization . |
| Outcome: | The proposed system achieves state-of-the-art results on WikiCoref, compared with a system trained on CoNLL. |
Lessons Learned from a Citizen Science Project for Natural Language Processing (2023.eacl-main)
Copied to clipboard
Jan-Christoph Klie, Ji-Ung Lee, Kevin Stowe, Gözde Şahin, Nafise Sadat Moosavi, Luke Bates, Dominic Petrak, Richard Eckart De Castilho, Iryna Gurevych
| Challenge: | Annotations are expensive and difficult to obtain, which is why many NLP systems outsource their work to paid crowdworkers. |
| Approach: | They propose to use Citizen Science to re-annotate parts of a pre-existing crowdsourced dataset to gain high-quality annotations. |
| Outcome: | The proposed approach yields high-quality annotations and motivated volunteers, but requires consideration of scalability, participation over time, and legal and ethical issues. |
Rolling the DICE on Idiomaticity: How LLMs Fail to Grasp Context (2025.acl-long)
Copied to clipboard
| Challenge: | Existing models fail to resolve idiomaticity when it depends on contextual understanding . idiom frequency influences performance but does not guarantee accurate interpretation. |
| Approach: | They propose a novel contrastive dataset to assess whether large language models can effectively leverage context to disambiguate idiomatic meanings. |
| Outcome: | The proposed model performs better on sentences deemed more likely by the model . collocational frequency and sentence probability influence performance but not accuracy . |
From Input Perception to Predictive Insight: Modeling Model Blind Spots Before They Become Errors (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for error or uncertainty estimation rely on logits, entropy, sampling variance. |
| Approach: | They propose an input-only method for anticipating language model errors using token-level likelihood features inspired by surprisal and the Uniform Information Density hypothesis. |
| Outcome: | The proposed method outperforms baseline models and standard models on linguistic datasets. |
Layer or Representation Space: What Makes BERT-based Evaluation Metrics Robust? (2022.coling-1)
Copied to clipboard
| Challenge: | Recent embedding-based evaluation metrics for text generation are based on measuring correlation with human evaluations on standard benchmarks. |
| Approach: | They examine the robustness of BERTScore, one of the most popular embedding-based metrics for text generation. |
| Outcome: | The embedding-based metrics that have 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. |