[8A]ハルシネーション・フェイク検出

13:30 〜 13:50

[8A-01] Multi-Hop Corpus for Detecting LLM Hallucinations

*Ali Ushtar¹、Lynden Steven²、的野 晃整²、天笠 俊之¹ (1. University of Tsukuba、2. AIST)

発表者区分：学生
論文種別：ショートペーパー
インタラクティブ発表：あり

キーワード：

Large Language Models.、Knowledge Graphs.、LLM Hallucination.

Knowledge graphs can reliably ground large language models (LLMs) in factual accuracy. Introducing KGs into multi-hop question answer (QA) can assess LLMs factual knowledge by testing their reasoning and inference skills. In this paper, we present a multi-hop QA dataset called EDQA, an entropy driven technique which generates the multi-hop questions from the KGs. Through experiments, we reveal that, our dataset effectively evaluates the factual accuracy of LLMs and detects hallucinations.