On Hallucination and Predictive Uncertainty in Conditional Language Generation
This addresses the problem of hallucinated facts in language generation models for NLP researchers, offering an incremental improvement through uncertainty analysis.
The paper investigates the relationship between hallucinations and predictive uncertainty in conditional language generation tasks like image captioning and data-to-text generation, finding that higher predictive uncertainty correlates with increased hallucination and that epistemic uncertainty is most indicative, leading to a beam search variant that reduces hallucination at the cost of some standard metric performance.
Despite improvements in performances on different natural language generation tasks, deep neural models are prone to hallucinating facts that are incorrect or nonexistent. Different hypotheses are proposed and examined separately for different tasks, but no systematic explanations are available across these tasks. In this study, we draw connections between hallucinations and predictive uncertainty in conditional language generation. We investigate their relationship in both image captioning and data-to-text generation and propose a simple extension to beam search to reduce hallucination. Our analysis shows that higher predictive uncertainty corresponds to a higher chance of hallucination. Epistemic uncertainty is more indicative of hallucination than aleatoric or total uncertainties. It helps to achieve better results of trading performance in standard metric for less hallucination with the proposed beam search variant.