Hallucinated Span Detection with Multi-View Attention Features
This addresses the specific issue of span-level hallucination detection, which is important for practical applications but has received less attention than output-level detection, representing an incremental advance in the field.
The study tackled the problem of detecting hallucinated spans in large language model outputs by extracting multi-view attention features and using a Transformer-based classifier, achieving improved performance over baselines on tasks with longer contexts like data-to-text and summarization.
This study addresses the problem of hallucinated span detection in the outputs of large language models. It has received less attention than output-level hallucination detection despite its practical importance. Prior work has shown that attentions often exhibit irregular patterns when hallucinations occur. Motivated by these findings, we extract features from the attention matrix that provide complementary views capturing (a) whether certain tokens are influential or ignored, (b) whether attention is biased toward specific subsets, and (c) whether a token is generated referring to a narrow or broad context, in the generation. These features are input to a Transformer-based classifier to conduct sequential labelling to identify hallucinated spans. Experimental results indicate that the proposed method outperforms strong baselines on hallucinated span detection with longer input contexts, such as data-to-text and summarisation tasks.