Survey of Hallucination in Natural Language Generation

This paper surveys the problem of hallucination in natural language generation (NLG), which has become more pressing as sequence-to-sequence and Transformer-based language models produce increasingly fluent and coherent text. The authors note that despite improvements in downstream tasks such as abstractive summarization, dialogue generation, and data-to-text generation, deep-learning generation is prone to hallucinating unintended text that degrades system performance and fails to meet user expectations. The survey is the first to review these efforts in a comprehensive manner.

The survey is organized into two parts: a general overview of hallucination metrics, mitigation methods, and future directions; and an overview of task-specific research progress in abstractive summarization, dialogue generation, generative question answering, data-to-text generation, and machine translation. By consolidating scattered work into one reference, it aims to facilitate collaborative efforts among researchers tackling hallucinated text in NLG.