Morocco

Recent advancements in Large Language Models (LLMs) have notably enhanced task-oriented dialogue systems, particularly in Dialogue State Tracking (DST), owing to their generative capabilities and strong generalization. Although recent approaches such as LDST and FnCTOD significantly improved cross-domain DST performance via supervised fine-tuning (SFT), these methods typically require substantial amounts of domain-specific data. In this paper, we address this limitation by employing Group Relative Policy Optimization (GRPO) - a critic-free reinforcement learning method that efficiently guides LLMs toward improved DST accuracy even under low-resource conditions. Our results on established DST benchmarks, including MultiWOZ 2.1 and 2.4, demonstrate that the RL approach achieves superior performance to existing methods while using significantly reduced out-of-domain training data. In addition, we found out that models pretrained specifically for tool-use tasks can be a better starting point, especially on small scales.

EACL 2026

Call, Reward, Repeat: Advancing Dialog State Tracking with GRPO and Function Calling

grpo

## 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*





<br>

EMNLP 2025

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

poster

#### *Message from the General Chair, Aline Villavicencio*
I’m delighted and honoured to welcome you to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), taking place in the beautiful city of Rabat, in Morocco, in March 24-29, 2026. EACL is the flagship European conference of the Association and EACL 2026 proudly continues our field’s tradition of excellence in scholarship, innovation, and inclusivity. I am deeply grateful to the many volunteers whose dedication, generosity, and tireless efforts have made this conference possible.
For the first time EACL is being hosted in the African continent. This is an important milestone for our community, and we are grateful to our Moroccan hosts for enabling this historic moment by bringing this edition of EACL to Rabat. We are also delighted that the Second Arabic NLP School is co-located with EACL. We hope attendees enjoy this wonderful opportunity to strengthen ties with the Computational Linguistics communities across the African continent. *[Read full message](https://drive.google.com/file/d/14NlmHvwM6fPJuMmOvVh7K0vtQbEyv3SZ/view?usp=sharing)*<br><br>

<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://underline.io/events/522/reception','_blank')">Go to Workshops and Tutorials Program</button></html>
<br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to EACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.eacl.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/eacl2026

Registration Required

Welcome to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL). EACL 2026 will be held in Rabat, Morocco, from March 24–29, 2026. 

Detecting personally identifiable information (PII) in user queries is critical for ensuring privacy in question-answering systems. Current approaches mainly redact all PII, disregarding the fact that some of them may be contextually relevant to the user’s question, resulting in a degradation of response quality. Large language models (LLMs) might be able to help determine which PII are relevant, but due to their closed source nature and lack of privacy guarantees, they are unsuitable for sensitive data processing. To achieve privacy-preserving PII detection, we propose CAPID, a practical approach that fine-tunes a locally owned small language model (SLM) that filters sensitive information before it is passed to LLMs for QA. However, existing datasets do not capture the context-dependent relevance of PII needed to train such a model effectively. To fill this gap, we propose a synthetic data generation pipeline that leverages LLMs to produce a diverse, domain-rich dataset spanning multiple PII types and relevance levels. Using this dataset, we fine-tune an SLM to detect PII spans, classify their types, and estimate contextual relevance. Our experiments show that relevance-aware PII detection with a fine-tuned SLM substantially outperforms existing baselines in span, relevance and type accuracy while preserving significantly higher downstream utility under anonymization.

CAPID: Context-Aware PII Detection for Question-Answering Systems

Music is a universal cultural practice that influences emotion, ritual and creativity, and it is now represented in many digital modalities: audio recordings, symbolic encodings (MIDI, MusicXML, ABC), visual scores and lyrics. Multimodal Large Language Models (MLLMs) have the ambition to process "everything", including music, and therefore promise to support musical analysis, creation and education. Despite this promise, systematic methods for evaluating whether a MLLM understands music are missing. Existing music‑focused benchmarks are fragmented, largely single‑modality, Western‑centric, and often do not require actual perception of the musical content; methodological details such as prompt design and answer‑extraction are frequently omitted or not discussed, and some evaluations rely on proprietary LLMs, hindering reproducibility and raising concerns about test‑data leakage. To fill this gap, this dissertation proposes to design a musically multimodal benchmark built on a transparent, fully open evaluation pipeline. The benchmark will present closed‑question‑answer items across four musical modalities, employ carefully engineered distractor options to enforce genuine perceptual engagement, and follow rigorously documented prompt‑selection and answer‑extraction procedures. It will further incorporate culturally diverse musical material beyond the dominant Western canon. Guided by three research questions: (1) how to devise robust, reproducible evaluation procedures, (2) how current MLLMs perform across modalities, and (3) how model scores relate to human musical abilities; the benchmark will enable precise diagnosis of model limitations, inform the development of more musically aware AI systems, and provide a principled basis for assessing practical usefulness to musicians and other stakeholders in the creative industry.

Thesis Proposal: Multimodal Benchmark for Music Understanding in Large Language Models

In-image machine translation is a sub-task of Image-Based Machine Translation that aims to substitute text embedded in images with its translation into another language. In the current work, we define a simple task with a synthetic dataset based on rendering parallel text over a plain background. Furthermore, we experiment with different optical character recognition, machine translation and image synthesis models to include in our ensemble. Then, we present our cascade approach as a pipeline that obtains the transcript of the original image, translates it, and generates a new image (image synthesis) similar to the original one. Finally, we compare the performance of our approach with several current state-of-the-art models, including an end-to-end approach, demonstrating its competitiveness.

In-Image Machine Translation. A Preliminary Modular Approach

This study proposes a method for learning subword correspondences in parallel sentence pairs using the EM algorithm. Conventional neural machine translation typically employs subword segmentation models trained. However, since existing methods do not consider parallel relationships, inconsistencies in word segmentation between source and target languages may hinder translation model training. Our approach leverages direct modeling of subword correspondences in parallel corpora, thereby improving segmentation consistency across languages. Experiments across multiple machine translation tasks confirm that our proposed method improves translation accuracy for many tasks.

Probabilistic Bilingual Subword Segmentation with Latent Subword Alignment

Lyrics translation must account for rhythm, rhyme, and singability in the translated lyrics. In this study, we focus on singability and investigate effective prompting methods for translating singable lyrics, including verification-guided and multi-round prompting, applied to large language models. First, we curate a multilingual lyrics translation dataset covering a total of six language directions across Chinese, Japanese, and English. Next, we evaluate seven prompting strategies, with instruction complexity increasing incrementally. The results show that multi-prompt strategies improve singability-related aspects, such as rhythmic alignment and phonological naturalness, compared to naive translation. Furthermore, human evaluations using songs created from translated lyrics suggest that moderately complex prompting strategies improve singable naturalness, while more complex strategies contribute to greater stability in perceived quality.

Towards Singable Lyrics Translation Using Large Language Models

Most vision-language models (VLMs) are trained on English-centric data, limiting their performance in other languages and cultural contexts. This restricts their usability for non-English-speaking users and hinders the development of multimodal systems that reflect diverse linguistic and cultural realities. In this work, we reproduce and adapt the LLaVA-Next methodology to create a set of Polish VLMs. We rely on a fully automated pipeline for translating and filtering existing multimodal datasets, and complement this with synthetic Polish data for OCR and culturally specific tasks. Despite relying almost entirely on automatic translation and minimal manual intervention, our approach yields strong results: we observe a +9.5 pp improvement over LLaVA‑1.6‑Vicuna‑13B on a Polish-adapted MMBench, along with higher-quality captions in generative evaluations, as measured by human annotators in terms of linguistic correctness. These findings highlight that large-scale automated translation, combined with lightweight filtering, can effectively bootstrap high-quality multimodal models for low-resource languages. Some challenges remain, particularly in cultural coverage and evaluation. To facilitate further research, we release our models, code, and datasets.

Annotation-Efficient Vision-Language Model Adaptation to the Polish Language Using the LLaVA Framework

This thesis investigates how polyvocal ontologies and Large Language Model (LLM) based Multi-Agent Systems (MAS) can operationalize perspective-aware knowledge extraction, preserving conflicting stakeholder interpretations as epistemically separable, queryable Knowledge Graphs (KGs). Current AI systems consolidate multiple perspectives into singular, decontextualized schemas, introducing representational bias and information loss. We propose a systematic framework addressing three interconnected research questions: (1) how to generate polyvocal ontology design patterns for high-stakes domains; (2) how to architect LLM-based MAS that extract perspective-conditioned facts while maintaining schema coherence and provenance traceability; and (3) whether such extractions achieve semantic diversity without sacrificing KG integrity. Evaluation is proposed on medical datasets, conducted with domain experts, to demonstrate the feasibility of perspective-aware extraction as a principled alternative to consensus-oriented KGs. Expected contributions include polyvocal ontology patterns, an ontology-orchestrated MAS extraction framework with auditable provenance, and empirical validation.

Thesis Proposal: A Multi-Agent System for Ontology-Based Perspective-Aware Knowledge Extraction

This paper introduces \textbf{DRAGOn}, method to design a RAG benchmark on a regularly updated corpus. It features recent reference datasets, a question generation framework, an automatic evaluation pipeline, and a public leaderboard.
Specified reference datasets allow for uniform comparison of RAG systems, while newly generated dataset versions mitigate data leakage and ensure that all models are evaluated on unseen, comparable data.
The pipeline for automatic question generation extracts the Knowledge Graph from the text corpus and produces multiple question-answer pairs utilizing modern LLM capabilities.
A set of diverse LLM-as-Judge metrics is provided for a comprehensive model evaluation.
We used Russian news outlets to form the datasets and demonstrate our methodology. We launch a public leaderboard to track the development of RAG systems and encourage community participation.

DRAGOn: Designing RAG On Periodically Updated Corpus

Writing style functions both as a vehicle of expression and as a marker of authorial identity. Stylometric methods enable automatic recognition of authors based on linguistic regularities, while recent advances in adversarial learning—demonstrate how data can be intentionally modified to prevent models from learning usable representations. Yet it remains unclear whether such perturbations, designed to disrupt machine learning processes, also influence human perception of style.

This thesis investigates how humans and models perceive writing style under controlled perturbations and whether manipulations that reduce algorithmic recognition likewise obscure stylistic identity for human readers. The study combines computational and behavioral approaches: constructing semantically controlled yet stylistically diverse text datasets, and conducting human evaluation experiments to compare recognition accuracy between models and readers.

The results are expected to clarify how linguistic cues contribute differently to human and algorithmic perception of style and to inform broader applications in authorship analysis, privacy-preserving text transformation, and creative expression. By situating writing style as a dimension of information quality, the research contributes to understanding how authenticity, anonymity, and expressivity interact in digital communication.

Thesis Proposal: Comparing Human and Model Perception of Writing Style under Controlled Perturbations

The use of large language models (LLMs) for generating responses to multiple-choice style questionnaires that were originally intended to be answered by humans is often a helpful or even necessary task, for example in persona simulation or during LLM alignment. Although the input and output versatility of generative LLMs is beneficial when adapting such questionnaires to machine use, it can be detrimental when mapping the generated text back to a closed set of possible answer options for evaluation or scoring. In this paper, we investigate the performance of smaller models for the classification of LLM outputs into the available answer options of multiple-choice questionnaires. We consider fine-tuned encoder-transformers as well as a rule-based approach on three datasets with differing answer option complexity. Surprisingly, we find that the best-performing neural approach still underperforms in comparison to our rule-based baseline, indicating that simple pattern-matching of answer options against LLM outputs might still be the most competitive solution for cleaning LLM responses to multiple-choice questionnaires.

Downloads

Next from EACL 2026

CAPID: Context-Aware PII Detection for Question-Answering Systems

Similar lecture

Beyond Single-User Dialogue: Assessing Multi-User Dialogue State Tracking Capabilities of Large Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES