Span-Level Hallucination Detection for LLM-Generated Answers

2025-04-25

AI Summary

This work addresses factual hallucination in large language model (LLM) text generation by proposing the first English–Arabic bilingual token-level hallucination detection framework. Methodologically, it innovatively integrates semantic role labeling (SRL) with retrieval-augmented textual entailment modeling and introduces a logit-driven token confidence calibration mechanism to enable interpretable span-level hallucination localization. Unlike conventional sentence-level binary classification, this framework achieves finer-grained detection with enhanced precision and attribution capability. Evaluated on the Mu-SHROOM benchmark, it establishes new state-of-the-art performance. Hallucinated spans are rigorously verified via cross-fact-checking using GPT-4 and LLaMA, significantly improving both detection accuracy and interpretability. The framework thus introduces a novel paradigm for multilingual, trustworthy evaluation of LLM outputs.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: Human-in-the-loop Machine Learning

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Detecting spans of hallucination in LLM-generated answers is crucial for improving factual consistency. This paper presents a span-level hallucination detection framework for the SemEval-2025 Shared Task, focusing on English and Arabic texts. Our approach integrates Semantic Role Labeling (SRL) to decompose the answer into atomic roles, which are then compared with a retrieved reference context obtained via question-based LLM prompting. Using a DeBERTa-based textual entailment model, we evaluate each role semantic alignment with the retrieved context. The entailment scores are further refined through token-level confidence measures derived from output logits, and the combined scores are used to detect hallucinated spans. Experiments on the Mu-SHROOM dataset demonstrate competitive performance. Additionally, hallucinated spans have been verified through fact-checking by prompting GPT-4 and LLaMA. Our findings contribute to improving hallucination detection in LLM-generated responses.

Problem

• Detecting hallucinated spans in LLM-generated answers
• Improving factual consistency via semantic role alignment
• Evaluating performance on English and Arabic texts

Innovation

• Semantic Role Labeling for answer decomposition
• DeBERTa-based entailment model for alignment
• Token-level confidence measures for refinement