Morocco

Fine-tuning Transformer models is often dominated by the backward computation in linear layers. In many NLP tasks, input sequences are short and padded to a fixed context length, inducing structured sparsity in the output gradients. We propose Sparsity-Exploiting Backward Pass (SEBP), a heuristic method that reduces backward computation by exploiting this sparsity with negligible memory overhead. We show that, for short input sequences, the output gradients of BERT-based and LLaMA models exhibit pronounced sparsity, allowing for optimisation in the backward computation. We optimized the autograd function in the linear layers, significantly reducing the number of FLOPs during the backward.

Our method achieves a backward pass speedup of approximately 2.15x for BERT-base on GLUE tasks and 1.99x for a 3B LLaMA model on reasoning benchmarks, while maintaining memory usage nearly identical to the regular PyTorch fine-tuning. Crucially, this speedup comes at no cost to performance. We show that our method matches standard convergence rates, offering a memory-efficient way to accelerate LLM fine-tuning.

EACL 2026

Acceleration of Backpropagation in Linear Layers of Transformer Models Based on Gradient Structure

memory efficiency.

gradient sparsity

backpropagation

large language models

efficient training

fine-tuning

transformers

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". 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 need to log in with the **email address you registered with**. 

ACL 2025

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 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)* 

<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>

To access this event page, you need to log in with the **email address you registered with**. 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.

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. 

Understanding sarcasm requires integrating cues from language, voice, and facial expression. Recent work has achieved impressive results using large multimodal Transformers, but such models are computationally expensive and often obscure how each modality contributes to the final prediction. This paper introduces a lightweight, interpretable framework for multimodal sarcasm detection that combines frozen text, audio, and visual embeddings from pretrained encoders through compact fusion heads. Using the MUStARD++Balanced dataset, we show that early fusion of textual and acoustic features improves over the best unimodal baseline. Character-specific evaluation further shows that sarcasm expressed through overt prosodic and visual cues is substantially easier to detect than monotone, context-dependent sarcasm. Additionally, we evaluate generalization to different characters through leave-one-speaker-out (LOSO) experiments and run ablation-style transfer experiments on two speakers with similar sarcasm distributions. These findings demonstrate that effective multimodal sarcasm understanding can emerge from frozen, resource-efficient representations without large-scale fine-tuning, emphasizing the importance of modality interaction and delivery style rather than model scale.

Scale Is All You Need 🙄: Analyzing Modality Interaction and Speaker Intent Without Fine-Tuning

Automatic story generation aims to produce coherent, engaging, and contextually consistent narratives with minimal or no human involvement, thereby advancing research in computational creativity and applications in human language technologies. The emergence of large language models has progressed the task, enabling systems to generate multi-thousand-word stories under diverse constraints. Despite these advances, maintaining narrative coherence, character consistency, storyline diversity, and plot controllability in generating stories is still challenging. In this survey, we conduct a systematic review of research published over the past four years to examine the major trends and key limitations in story generation methods, model architectures, datasets, and evaluation methodologies. Based on this analysis of 57 included papers, we propose developing new evaluation metrics and creating more suitable datasets, together with ongoing improvement of narrative coherence and consistency, as well as their exploration in practical applications of story generation, as actions to support continued progress in automatic story generation.

Text-to-Text Automatic Story Generation: A Survey

Indian languages represent a highly multilingual and low-resource speech ecosystem, where the scarcity of high-quality parallel speech corpora significantly limits the development of speech-to-speech translation systems. Most existing approaches rely on cascaded pipelines that combine automatic speech recognition (ASR), machine translation (MT), and text-to-speech synthesis (TTS). While effective, these cascaded systems often suffer from cumulative error propagation, increased latency, and higher computational complexity, particularly in low-resource Indian languages. To address these challenges, my doctoral work proposes a novel sequence-to-sequence direct speech translation framework capable of translating speech from one Indian language to another without relying on intermediate text representations. Recent advances in deep learning, however, indicate that direct speech translation architectures can surpass conventional cascaded systems in both efficiency and translation quality, motivating the design of our fully end-to-end solution. We aim to release an initial dataset comprising at least 120,000 real speech samples within a 6–12 month timeframe.

Thesis Proposal: Development of End-to-End Speech Translation Models for Indian Languages

Recent advances in Sparse Autoencoders (SAEs) have revealed interpretable features within large language models (LLMs), including features that are specific to individual languages.
In prior work, these features have been used to steer a model’s output language.
However, the impact of SAE-based language steering on output quality and task performance, as well as its relationship to simpler prompting-based approaches, remains unclear.
In this work, we study the effects of language steering using SAE features across multiple tasks and models.
We apply language-specific SAE feature steering to three LLMs from two model families and evaluate it on a translation task and a multilingual question-answering task.
We compare SAE-based steering against prompting and language neuron-based steering, and examine a combined prompting-and-steering approach.
On the translation task, SAE feature steering achieves an average target-language accuracy of 92\% across models and languages, consistently outperforming language neuron-based steering, but slightly underperforming prompting in language accuracy and output quality.
In contrast, on the multilingual question-answering task, SAE-based steering enables stronger language control than prompting, and combining steering with prompting yields the best overall language control and task performance.
These findings demonstrate the potential of SAE features as a tool for controllable multilingual generation.

Evaluating the Impact of SAE-based Language Steering on LLM Performance

Police interrogation transcripts are key evidential documents, yet their linguistic form is rarely systematically analyzed, despite directly shaping judicial interpretation. This study presents the first computational forensic linguistic profiling of Italian police transcripts, focusing on the two transcription formats used in practice: narrative monologues and question-answer (Q-A) transcripts. Using automated extraction of 147 linguistic features, we analyze 50 authentic transcripts against a multi-genre Italian reference corpus to support more transparent evaluation of police transcripts by clarifying how transcription formats systematically shape evidential interpretation in judicial contexts. Narrative monologues exhibit deeper syntactic embedding, higher past-tense usage, and more first-person singular verbs, supporting coherent and temporally ordered recounting of events. Q-A transcripts, by contrast, show longer subordinate chains, more clausal complements, and higher pronoun frequency, reflecting interactive turn-taking and procedural dynamics. Rather than aiming at predictive classification, the study reveals the linguistic mechanisms shaping transcription formats and demonstrates that structurally and legally informed features reliably distinguish them. Computational models reliably capture genre-specific cues, offering scalable, empirically grounded insights into transcription practices and evidential reliability.

A Computational Forensic Linguistic Analysis of Narrative and Question-Answer Structures in Italian Police Interrogation Transcripts

The spread of misinformation has prompted extensive research on machine-learning–based fake news detection. However, existing datasets differ substantially in content distributions and annotation policies, complicating fair evaluation and generalization assessment. We refer to these structural differences as dataset bias. In this study, we quantitatively analyze dataset bias across multiple public fake news datasets (Kaggle, FNN, ISOT, and NELA-GT-2019/2020) with different annotation granularities, including article-level and publisher-level labels. Using document embedding–based similarity analysis and article category distributions, we examine how such biases affect detection performance under in-dataset and cross-dataset evaluation settings. Furthermore, to leverage large-scale but coarse-grained publisher-level data, we compare proxy-label training with a semi-supervised learning approach based on Virtual Adversarial Training (VAT). Our results show that detection performance strongly depends on dataset-specific biases, and that proxy-label training and SSL exhibit complementary, and sometimes opposite, strengths depending on whether the evaluation emphasizes in-dataset performance or cross-dataset generalization. These findings highlight the importance of appropriate training strategies and evaluation protocols when using heterogeneous fake news datasets.

Fake News Detection Strategies under Dataset Bias: Using Large-scale Coarse-grained Labels

Training a language model for low-resource languages is challenging due to data scarcity and computational cost. Tokenizer transfer offers a way to adapt a pre-trained model to a new tokenizer without full retraining, improving efficiency and cross-lingual applicability. To the best our of knowledge, we present the first controlled evaluation of tokenizer transfer on monolingually pretrained base models trained on language-specific corpora, Orthogonal Mapping Pursuit (OMP) and Fast Vocabulary Transfer (FVT), across six languages and multiple finetuning regimes. Using the Goldfish model family, we evaluate using byte-normalized log-perplexity and MultiBlimp accuracy for target-language adaptability, source-language retention, and the interaction between transfer and monolingual or mixed finetuning. OMP with monolingual target finetuning yields the best target-language scores (lower log-perplexity and higher MultiBlimp) among our evaluated conditions, compared with (i) a model trained only on the source language, (ii) a model trained on a smaller amount of target-language data, and (iii) the source language model adapted via standard finetuning on the target data. The results suggest tokenizer transfer is a compute-efficient alternative for low-resource LM training: train a monolingual tokenizer for the target language, transfer it to a larger pre-trained model, and fine-tune using the target data.

Efficient Low-Resource Language Models Using Tokenizer Transfer

Nested named entity recognition identifies entities contained within other entities, but requires expensive multi-level annotation. While flat NER corpora exist abundantly, nested resources remain scarce. We investigate whether models can learn nested structure from flat annotations alone, evaluating four approaches: string inclusions (substring matching), entity corruption (pseudo-nested data), flat neutralization (reducing false negative signal), and a hybrid fine-tuned + LLM pipeline. On NEREL, a Russian benchmark with 29 entity types where 21\% of entities are nested, our best combined method achieves 26.37\% inner F1, closing 40\% of the gap to full nested supervision. Code is available at https://github.com/fulstock/Learning-from-Flat-Annotations.

Learning Nested Named Entity Recognition from Flat Annotations

This paper investigates the capabilities of LLMs to detect and explain fine-grained emotional social influence techniques in textual dialogues, as well as human preferences for technique explanations. We present findings from our two studies. In Study 1, a dataset of 238 Polish dialogues is introduced, each annotated with detailed span-level labels. On this data, we evaluate the performance of LLMs on two tasks: detecting 11 emotional social influence techniques and identifying text spans corresponding to specific techniques. The results indicate that current LLMs demonstrate limited effectiveness in accurately detecting fine-grained emotional social influence.
In Study 2, we examine various LLM-generated explanations through human pairwise preferences and four criteria: comprehensibility, cognitive coherence, completeness, and soundness, with the latter two emerging as the most influential on general human preference. All data, including human annotations, are publicly available as the EmoSocInflu dataset (https://github.com/social-influence/emo-soc-influ). 
Our findings highlight a critical need for further advancement in the field. As LLM-supported manipulation grows, it is essential to promote public understanding of social influence mechanisms, enabling individuals to critically recognize and interpret the subtle forms of manipulation that shape public opinion.

From Detection to Explanation: Modeling Fine-Grained Emotional Social Influence Techniques with LLMs and Human Preferences

Despite their recent success, the geospatial reasoning capabilities of large language models (LLMs)—which require understanding spatial relationships among real-world geo-entities—remain underexplored.
We propose an automatic method for constructing compositional geographic question answering datasets that jointly consider spatial and entity constraints.
The generated dataset serves as a principled benchmark for evaluating how LLMs coordinate spatial computation with entity-level understanding under diverse compositional settings.
We evaluate two state-of-the-art LLMs, GPT-5.2 and Gemini 3 Flash, on our dataset. Experimental results show that while the models perform relatively well on questions involving rich entity grounding, their accuracy drops substantially on questions requiring precise quantitative spatial reasoning, such as distance estimation and containment judgment.
Our dataset is publicly available for research and reproduction.

Downloads

Next from EACL 2026

Scale Is All You Need 🙄: Analyzing Modality Interaction and Speaker Intent Without Fine-Tuning

Similar lecture

Forward Knows Efficient Backward Path: Saliency-Guided Memory-Efficient Fine-tuning of Large Language Models