Austria

The Multilingual shared-task on Hallucinations and Related Observable Overgeneration Mistakes in the SemEval-2025 competition aims to detect hallucination spans in the outputs of instruction-tuned LLMs in a multilingual con-
text. In this paper, we address the detection of span hallucinations by applying an ensemble of approaches. In particular, we synthesized a dataset and fine-tuned LLM to detect hallucination spans. In addition, we combined this approach with a white-box method based on uncertainty quantification techniques. Using our combined pipeline, we achieved 3rd place in detecting span hallucinations in Arabic, Catalan, Finnish, Italian, and ranked within the top ten for the rest of the languages

ACL 2025

SmurfCat at SemEval-2025 Task 3: Bridging External Knowledge and Model Uncertainty for Enhanced Hallucination Detection

### 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**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

workshop paper

This paper presents the method for the unlearning of sensitive information from
large language models as applied in the SemEval 2025 Task 4 challenge. The unlearning pipeline consists of two phases. In phase I, the model is instructed to forget specific datasets, and in phase II, the model is stabilized using a retention dataset. Unlearning with these methods secured a final score of 0.420 with the 2nd honorary mention in the 7B parameter challenge and a score of 0.36 in the 13th position for the 1B parameter challenge. The paper presents a background study, a brief literature review, and a gap analysis, as well as the methodology employed in our work titled NeuroReset. The training methodology and evaluation metrics are also presented, and the trade-offs between unlearning efficiency and model performance are discussed. The contributions of the paper are systematic unlearning, a comparative analysis of unlearning methods, and an empirical analysis of model performance post-unlearning.

NeuroReset : LLM Unlearning via Dual Phase Mixed Methodology

Hallucinations are one of the major problems of LLMs, hindering their trustworthiness and deployment to wider use cases. However, most of the research on hallucinations focuses on English data, neglecting the multilingual nature of LLMs. This paper describes our submission to the" SemEval-2025 Task-3 — Mu-SHROOM, the Multilingual Shared-task on Hallucinations and Related Observable Overgeneration Mistakes". We propose a two-part pipeline that combines retrieval-based fact verification against Wikipedia with a BERT-based system fine-tuned to identify common hallucination patterns. Our system achieves competitive results across all languages, reaching top-10 results in eight languages, including English. Moreover, it supports multiple languages beyond the fourteen covered by the shared task. This multilingual hallucination identifier can help to improve LLM outputs and their usefulness in the future.

TUM-MiKaNi at SemEval-2025 Task 3: Towards Multilingual and Knowledge-Aware Non-factual Hallucination Identification

The Mu-SHROOM competition in the SemEval-2025 Task 3 aims to tackle the problem of detecting spans with hallucinations in texts, generated by Large Language Models (LLMs). Our developed system, submitted to this task, is a joint architecture that utilises Named Entity Recognition (NER), Retrieval-Augmented Generation (RAG) and LLMs to gather, compare and analyse information in the texts provided by organizers. We extract entities potentially capable of containing hallucinations with NER, aggregate relevant topics for them using RAG, then verify and provide a verdict on the extracted information using the LLMs. This approach allowed with a certain level of quality to find hallucinations not only in facts, but misspellings in names and titles, which was not always accepted by human annotators in ground truth markup. We also point out some inconsistencies within annotators spans, that perhaps affected scores of all participants.

Advacheck at SemEval-2025 Task 3: Combining NER and RAG to Spot Hallucinations in LLM Answers

Understanding idioms in multimodal contexts poses significant challenges due to data scarcity, idiomatic ambiguity, and the need for effective alignment of visual and textual inputs. In this work, we introduce MIRA (Multimodal Idiom Recognition and Alignment), a training- free framework designed to address these challenges on the SemEval-2025 Task 1 (AdMIRe) benchmark. MIRA leverages powerful closed- source large language models (LLMs) and integrates three key innovations: bias correction via in-context learning, multi-step semantic- visual fusion, and a self-revision mechanism that iteratively refines its outputs through backward verification. By systematically processing and fusing multimodal inputs, MIRA generates high-quality, fine-grained image-text repre- sentations that enhance idiom comprehension across different languages and cultural contexts. Experimental evaluations in both English and Portuguese demonstrate that our approach achieves robust performance without the need for additional training, setting a new standard for multimodal idiom recognition.

PALI-NLP at SemEval 2025 Task 1: AdMIRe Advancing Multimodal Idiomaticity Representation

Question answering using Large Language Models has gained significant popularity in both everyday communication and at the workplace. However, certain tasks, such as querying tables, still pose challenges for commercial and open-source chatbots powered by advanced deep learning models. Addressing these challenges requires specialized approaches. 
During the SemEval-2025 Task 8 competition focused on tabular data, our solution achieved 86.21%accuracy and took 2nd place out of 100 teams. In this paper we present ten methods that significantly improve the baseline solution.

Samsung Research Poland at SemEval-2025 Task 8: LLM ensemble methods for QA over tabular data

Food contamination and associated illnesses represent significant global health challenges, leading to thousands of deaths worldwide. As the volume of food-related incident reports on web platforms continues to grow, there is a pressing demand for systems capable of detecting food hazards effectively. Furthermore, explainability in food risk detection is crucial for building trust in automated systems, allowing humans to validate predictions. SemEval-2025 Task 9 proposes a food hazard detection challenge to address this issue, utilizing content extracted from websites. This task is divided into two sub-tasks. Sub-task 1 involves classifying the type of hazard and product, while sub-task 2 focuses on identifying precise hazard and product ``vectors'' to offer detailed explanations for the predictions. This paper presents our participation in this task, where we introduce a transformer-based method. We fine-tune an enhanced version of the BERT transformer to process lengthy food incident reports. Additionally, we combine the transformer’s contextual embeddings to enhance its contextual representation for hazard and product ``vectors'' prediction. The experimental results reveal the competitive performance of our proposed method in this task, which achieved 7th place in both sub-tasks. We have released our code at https://github.com/AhmadMonir-CSECU/SemEval-2025_Task9.

CSECU-Learners at SemEval-2025 Task 9: Enhancing Transformer Model for Explainable Food Hazard Detection in Text

Multilingual hallucination detection stands as an underexplored challenge, which the Mu-SHROOM shared task seeks to address. In this work, we propose an efficient, training-free LLM prompting strategy that enhances detection by translating multilingual text spans into English. Our approach achieves competitive rankings across multiple languages, securing two first positions in low-resource languages.
The consistency of our results highlights the effectiveness of our translation strategy for hallucination detection, demonstrating its applicability regardless of the source language.

AILS-NTUA at SemEval-2025 Task 3: Leveraging Large Language Models and Translation Strategies for Multilingual Hallucination Detection

We present our submission to SemEval 2025 Task 8: Question Answering on Tabular Data, which challenges participants to develop systems capable of answering natural language questions on real-world tabular datasets. Our approach aims at generating Pandas code that can be run on such datasets to produce the desired answer. 
The approach consists in fine-tuning a Small Language Model (SLM) through Preference Optimization on both positive and negative examples generated by a teacher model. A base SLM is first elicited to produce the code to compute the answer to a question through a Chain of Thought (CoT) prompt. We performed extensive testing on the DataBench development set, exploring a variety of prompts, eventually settling on a detailed instruction prompt, followed by two-shot examples. 
Due to hardware constraints, the base model was an SLM with ≤ 8 billion parameters. We then fine-tuned the model through Odds Ratio Preference Optimization (ORPO) using as training data the code produced by a teacher model on the DataBench training set. The teacher model was GPT-4o, whose code was labeled preferred, while the code generated by the base model was rejected. This increased the accuracy on the development set from 71% to 85%.
Our method demonstrated robust performance in answering complex questions across diverse datasets, highlighting the effectiveness of combining small LLMs with supervised fine-tuning and automated code execution for tabular question answering.

Dataground at SemEval-2025 Task 8: Small LLMs and Preference Optimization for Tabular QA

In today's digital era, individuals convey their feelings, viewpoints, and perspectives across various platforms in nuanced and intricate ways. At times, these expressions can be challenging to articulate and interpret. Emotion recognition aims to identify the most relevant emotions in a text that accurately represent the author's psychological state. Despite its substantial impact on natural language processing (NLP), this task has primarily been researched only in high-resource languages. To bridge this gap, SemEval-2025 Task 11 introduces a multilingual emotion recognition challenge encompassing 32 languages, promoting broader linguistic inclusivity in emotion recognition. This paper presents our participation in this task, where we introduce a language-specific fine-tuned transformer-based system for emotion recognition and emotion intensity prediction. To enhance generalization, we incorporate a multi-sample dropout strategy. Our approach is evaluated across 11 languages, and experimental results demonstrate its competitive performance, achieving top-tier results in certain languages.

CSECU-Learners at SemEval-2025 Task 11: Multilingual Emotion Recognition and Intensity Prediction with Language-tuned Transformers and Multi-sample Dropout

Food safety incidents cause serious threats to public health, requiring efficient detection systems. This study contributes to SemEval 2025 Task 9: Food Hazard Detection by leveraging insights from existing literature and using multiple BERT-based models for multi-label classification of food hazards and product categories. Using a dataset of food recall notifications, we applied preprocessing techniques
to prepare data and address challenges like class imbalance. Experimental results show strong hazard classification performance on ensembled models such as DistilBERT, SciBERT, and DeBERTa but highlight product classification variability. Building on Tyagi et al. (Tyagi et al., 2023) and Madry et al.’s (Rebuffi et al., 2021) work, we explored strategies like ensemble modeling and data augmentation to improve accuracy and explainability, paving the way for scalable food safety solutions.

Downloads

Next from ACL 2025

NeuroReset : LLM Unlearning via Dual Phase Mixed Methodology

Similar lecture

TUM-MiKaNi at SemEval-2025 Task 3: Towards Multilingual and Knowledge-Aware Non-factual Hallucination Identification

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2025

NeuroReset : LLM Unlearning via Dual Phase Mixed Methodology

Similar lecture

TUM-MiKaNi at SemEval-2025 Task 3: Towards Multilingual and Knowledge-Aware Non-factual Hallucination Identification

Downloads