Automated Detection of Factual Hallucinations in Large Language Model Outputs: A Deep Learning Framework for Span-Level Classification
Abhishek Pipliya, Pushpendra Kumar Mishra, Jaya Dubey, Shruti Lakhotiya, Mahima Jain, Ankita Singh Bhadoriya
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Large language models (LLMs) such as GPT-4, LLaMA-2, and Gemini have demonstrated remarkable capabilities in natural language generation. However, they remain critically prone to producing factually incorrect or unverifiable content, a phenomenon widely known as hallucination which poses significant risks in high-stakes domains such as healthcare, legal systems, and education. Existing hallucination detection methods primarily perform binary, sentence-level classification, identifying whether an output is hallucinated or not, but fail to localize specific hallucinated tokens or phrases within the generated text. This coarse granularity limits their practical utility in real-world deployment. This paper proposes HalluDetect, a hybrid deep learning framework that combines BERT contextual embeddings with a Bidirectional Long Short-Term Memory (BiLSTM) network for fine-grained, span-level hallucination detection, formulated as a token sequence labeling task. Each token in the LLM output is classified as HAL (hallucinated) or NON-HAL, enabling precise localization of factual errors. The proposed model is trained and evaluated on three publicly available benchmark datasets: HaluEval (35,000 samples), TruthfulQA (817 questions), and Mu-SHROOM (SemEval 2025 multilingual span annotations). Experimental results demonstrate that HalluDetect achieves 85.7% token-level F1- score and 64.2% span-level Intersection over Union (IoU), outperforming BERT, DeBERTa-v3, RoBERTa-large, and semantic entropy baselines by 8.5-16.1% in F1 and 25.8% in span IoU. Cross-dataset evaluation confirms strong generalization across question answering, dialogue, and summarization domains.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
计算机 / AITopic Modeling
Misinformation and Its Impacts · Natural Language Processing Techniques