China

Scientific visual question answering poses significant challenges for vision-language models due to the complexity of scientific figures and their multimodal context. Traditional approaches treat the figure and accompanying text (e.g., questions and answer options) as separate inputs. EXAMS-V introduced a new paradigm by embedding both visual and textual content into a single image. However, even state-of-the-art proprietary models perform poorly on this setup in zero-shot settings, underscoring the need for task-specific fine-tuning. To address the scarcity of training data in this &quot;text-in-image&quot; format, we synthesize a new dataset by converting existing separate image-text pairs into unified images. Fine-tuning a small multilingual multimodal model on a mix of our synthetic data and EXAMS-V yields notable gains across 13 languages, demonstrating strong average improvements and cross-lingual transfer.

EMNLP 2025

A Simple Data Augmentation Strategy for Text-in-Image Scientific VQA

Scientific visual question answering poses significant challenges for vision-language models due to the complexity of scientific figures and their multimodal context. Traditional approaches treat the figure and accompanying text (e.g., questions and answer options) as separate inputs. EXAMS-V introduced a new paradigm by embedding both visual and textual content into a single image. However, even state-of-the-art proprietary models perform poorly on this setup in zero-shot settings, underscoring the need for task-specific fine-tuning. To address the scarcity of training data in this "text-in-image" format, we synthesize a new dataset by converting existing separate image-text pairs into unified images. Fine-tuning a small multilingual multimodal model on a mix of our synthetic data and EXAMS-V yields notable gains across 13 languages, demonstrating strong average improvements and cross-lingual transfer.

workshop paper

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community."<br>

*Message from the General Chair, Dirk Hovy*

[**Link to Conference Handbook**](https://drive.google.com/file/d/1johU5QqVVYO4RfH7QcIORr7qrVBdzdwC/view?usp=sharing)





<br>

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Knowledge distillation (KD) is a popular method of transferring knowledge from a large "teacher" model to a small "student" model. Previous work has explored various layer-selection strategies (e.g., forward matching and in-order random matching) for intermediate-layer matching in KD, where a student layer is forced to resemble a certain teacher layer. In this work, we revisit such layer-selection strategies and observe an intriguing phenomenon that layer-selection strategy does not matter (much) in intermediate-layer matching---even seemingly nonsensical matching strategies such as reverse matching still result in surprisingly good student performance. We provide an interpretation for this phenomenon by examining the angles between teacher layers viewed from the student's perspective. Our work sheds light on KD practice, as layer-selection strategies may not be the main focus of KD system design and vanilla forward matching works well in most setups.

Revisiting Intermediate-Layer Matching in Knowledge Distillation: Layer-Selection Strategy Doesn’t Matter (Much)

Large language models (LLMs) excel in generating fluent utterances but can lack reliable grounding in verified information. At the same time, knowledge-graph-based fact-checkers deliver precise and interpretable evidence, yet suffer from limited coverage or latency. By integrating LLMs with knowledge graphs and real-time search agents, we introduce a hybrid fact-checking approach that leverages the individual strengths of each component. Our system comprises three autonomous steps: 1) a Knowledge Graph (KG) Retrieval for rapid one‑hop lookups in DBpedia, 2) an LM-based classification guided by a task-specific labeling prompt, producing outputs with internal rule-based logic, and 3) a Web Search Agent invoked only when KG coverage is insufficient. Our pipeline achieves an F1 score of 0.93 on the FEVER benchmark on the Supported/Refuted split without task‑specific fine‑tuning. To address Not enough information cases, we conduct a targeted reannotation study showing that our approach frequently uncovers valid evidence for claims originally labeled as Not Enough Information (NEI), as confirmed by both expert annotators and LLM reviewers. With this paper, we present a modular, open-source fact-checking pipeline with fallback strategies and generalization across datasets.

Hybrid Fact-Checking that Integrates Knowledge Graphs, Large Language Models, and Search-Based Retrieval Agents Improves Interpretable Claim Verification

Warning: This paper contains explicit statements of offensive stereotypes which may be upsetting.

Stereotypes vary across cultural contexts, making it essential to understand how language models encode social biases. MultiLingualCrowsPairs is a dataset of culturally adapted stereotypical and anti-stereotypical sentence pairs across nine languages. While prior work has primarily reported average fairness metrics on masked language models, this paper analyzes social biases in generative models by disaggregating results across specific bias types.

We find that although most languages show an overall preference for stereotypical sentences, this masks substantial variation across different types of bias, such as gender, religion, and socioeconomic status. Our findings underscore that relying solely on aggregated metrics can obscure important patterns, and that fine-grained, bias-specific analysis is critical for meaningful fairness evaluation.

Insights from a Disaggregated Analysis of Kinds of Biases in a Multicultural Dataset

As Large Language Models (LLMs) gain mainstream public usage, understanding how users interact with them becomes increasingly important. Limited variety in training data raises concerns about LLM reliability across different language inputs. To explore this, we test several LLMs using functionally equivalent prompts expressed in different English sublanguages. We frame this analysis using Question-Answer (QA) pairs, which allow us to detect and evaluate appropriate and anomalous model behavior. We contribute a cross-LLM testing method and a new QA dataset translated into AAVE and WAPE variants. Early results reveal a notable drop in accuracy for one sublanguage relative to the baseline.

That Ain't Right: Assessing LLM Performance on QA in African American and West African English Dialects

News classification is a key task in Natural Language Processing (NLP) that involves the automatic categorization of articles into predefined topics. While extensively studied in high-resource languages such as English, low-resource languages such as Amharic face significant challenges owing to limited datasets and the absence of language-specific model adaptations.
This study contributes to Amharic news classification by expanding the existing Amharic news dataset from 50,000 to 144,000 articles, nearly tripling its size. We evaluate five transformer-based models—mBERT, XLM-R, DistilBERT, AfriBERTa, and AfroXLM—on this enriched dataset to improve classification accuracy and effectiveness.
The model performance was assessed using accuracy, precision, recall, and F1-score, with careful attention to computational efficiency and overfitting. Among the models, AfriBERTa and XLM-R achieved the highest F1-scores of 90.25\% and 90.11\%, respectively, significantly surpassing the Naïve Bayes baseline, which reached only 60.3\% accuracy on the former dataset. These findings underscore the importance of both advanced transformer architectures and larger, high-quality datasets in boosting classification performance for under-resourced languages such as Amharic, offering valuable insights and resources for future NLP research.

Amharic News Topic Classification: Dataset and Transformer-Based Model Benchmarks

Sarcasm is a complex linguistic and pragmatic phenomenon where expressions convey meanings that contrast with their literal interpretations, requiring sensitivity to the speaker's intent and context. Misinterpreting sarcasm in collaborative human–AI settings can lead to under- or overreliance on LLM outputs, with consequences ranging from breakdowns in communication to critical safety failures. We introduce Sarc7, a benchmark for fine-grained sarcasm evaluation based on the MUStARD dataset, annotated with seven pragmatically defined sarcasm types: self-deprecating, brooding, deadpan, polite, obnoxious, raging, and manic. These categories are adapted from prior linguistic work and used to create a structured dataset suitable for LLM evaluation. For classification, we evaluate multiple prompting strategies—zero-shot, few-shot, chain-of-thought (CoT), and a novel emotion-based technique—across five major LLMs. Emotion-based prompting yields the highest macro-averaged F1 score of 0.3664 (Gemini 2.5), outperforming CoT for several models and demonstrating its effectiveness in sarcasm type recognition. For sarcasm generation, we design structured prompts using fixed values across four sarcasm-relevant dimensions: incongruity, shock value, context dependency, and emotion. Using Claude 3.5 Sonnet, this approach produces more subtype-aligned outputs, with human evaluators preferring emotion-based generations 38.46% more often than zero-shot baselines. Sarc7 offers a foundation for evaluating nuanced sarcasm understanding and controllable generation in LLMs, pushing beyond binary classification toward interpretable, emotion-informed language modeling.

Sarc7: Evaluating Sarcasm Detection and Generation with Seven Types and Emotion-Informed Techniques

Recent advances in Large Language Models (LLMs) have enhanced the fluency and coherence of Conversational Recommendation Systems (CRSs), yet emotional intelligence remains a critical gap. In this study, we systematically evaluate the emotional behavior of six state-of-the-art LLMs in CRS settings using the ReDial and INSPIRED datasets. We propose an emotion-aware evaluation framework incorporating metrics such as Emotion Alignment, Emotion Flatness, and per-emotion F1-scores. Our analysis shows that most models frequently default to emotionally flat or mismatched responses, often misaligning with user affect (e.g., joy misread as neutral). We further examine patterns of emotional misalignment and their impact on user-centric qualities such as personalization, justification, and satisfaction. Through qualitative analysis, we demonstrate that emotionally aligned responses enhance user experience, while misalignments lead to loss of trust and relevance. This work highlights the need for emotion-aware design in CRS and provides actionable insights for improving affective sensitivity in LLM-generated recommendations.

Emotionally Aware or Tone-Deaf? Evaluating Emotional Alignment in LLM-Based Conversational Recommendation Systems

Jailbreaking, the phenomenon where specific prompts cause LLMs to assist with harmful requests, remains a critical challenge in NLP, particularly in non-English and lower-resourced languages. To address this, we introduce MULBERE, a method that extends the method of Targeted Latent Adversarial Training (T-LAT) to a multilingual context. We first create and share a multilingual jailbreak dataset spanning high-, medium-, and low-resource languages, and then fine-tune LLaMA-2-7b-chat with interleaved T-LAT for jailbreak robustness and chat examples for model performance. Our evaluations show that MULBERE reduces average multilingual jailbreak success rates by 75\% compared to the base LLaMA safety training and 71\% compared to English-only T-LAT while maintaining or improving standard LLM performance.

MULBERE: Multilingual Jailbreak Robustness Using Targeted Latent Adversarial Training

Language models perpetuate dialect bias, associating African American English (AAE) with negative traits and outcomes. We propose JustDial, a lightweight finetuning framework aligning character trait associations between meaning-matched AAE and Standardized American English (SAE) text, while preserving general model fluency through a KL-divergence regularization term. Experiments on GPT2-Medium show that JustDial successfully removed any statistically significant correlation between dialect and predicted occupational prestige and reduced conviction and death-sentencing disparities by more than 98.7\%, with only 100,000 text examples and one epoch of LoRA finetuning. Though this debiasing comes at the cost of general model performance, adjusting the regularization term in JustDial enables a navigable debiasing-performance tradeoff space. JustDial provides the first proof-of-concept towards mitigating dialect prejudice in language models.

JustDial: Language Model Dialect Debiasing Using Biased Character Trait Associations

Multilingual Pre-trained Language Models (multiPLMs) trained with the Masked Language Modeling (MLM) objective exhibit suboptimal performance on cross-lingual downstream tasks for Low-Resource Languages (LRLs). Continually pre-training these multiPLMs with the Translation Language Modeling (TLM) objective on parallel data improves the cross-lingual performance. However, both MLM and TLM mask tokens randomly, which does not guarantee optimal representation learning. In this paper, we introduce a novel masking strategy, Linguistic Entity Masking (LEM) to improve the cross-lingual representations of existing multiPLMs. In contrast to MLM and TLM, LEM limits masking to the linguistic entities nouns, verbs and Named Entities, which hold a higher prominence in a sentence. We hypothesise that masking linguistically significant linguistic entities should contribute to effective representation learning. Empirically, we prove this using two downstream tasks with three LRL pairs, English-Sinhala, English-Tamil, and Sinhala-Tamil, and show that our LEM-based learning returns superior results compared to MLM+TLM.

Downloads

Next from EMNLP 2025

Revisiting Intermediate-Layer Matching in Knowledge Distillation: Layer-Selection Strategy Doesn’t Matter (Much)

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES