Reproduce “A Token-level Reference-free Hallucination Detection Benchmark for Free-form Text Generation” on OpenResearch

About this paper

Tianyu Liu, Yizhe Zhang, Chris Brockett, Yi Mao, Zhifang Sui, Weizhu Chen, Bill Dolan

Researchers from Peking University, Meta AI, and Microsoft Corporation established a token-level, reference-free hallucination detection benchmark for free-form text generation and introduced the HADES dataset. Their baseline transformer models achieved approximately 71-73% accuracy in identifying hallucinations at the token level, providing a more granular approach to factual consistency checks in generative AI outputs.

Key insights

• Contextualized transformer models are crucial for effective fine-grained hallucination detection, significantly outperforming traditional feature-based methods.
• Access to bidirectional context (offline setting) consistently improves detection performance compared to unidirectional context (online setting), highlighting the value of full textual information.
• Hallucinated tokens often exhibit higher word entropy, and counter-intuitively, sometimes higher average word probabilities from the generative model, suggesting overconfident erroneous outputs.
• Named entities are particularly susceptible to hallucination, with over 90% of identified named entity spans being labeled as hallucinatory in the HADES dataset.

Summary by alphaXiv.