A Text Generation Hallucination Detection Frame-work Based on Fact and Semantic Consistency
Xinyu Li, Yongbing Gao, Weihao Li, Lidong Yang · 2024
With the development of pre-trained language mod-els, significant progress has been made in text generation. How-ever, these models often face inconsistencies with facts and contra-dictions in sentences, known as the “hallucination” problem of language models. To address this issue, HalluCheckNet, a mechanism thatintegrates factual and semantic consistency for hallucination detection, is proposed. By combining pre-trained models with the Plug-and-Play Language Model (PPLM) framework, HalluCheck-Net treats the degree of hallucination as a controllable attribute, using controlled text generation techniques to effectively reduce model hallucinations. Experiments conducted on GPT2 have achieved high-quality generation of social media, review, and news texts. The results demonstrate that this method not only maintains text fluency but also significantly reduces factual and semantic er-rors, making the generated texts more coherent and credible.