Towards understanding and mitigating the hallucinations in NLP and Speech

Abstract

With the recent advances in natural language processing, thanks to deep learning architectures such as the transformer, the performances on many of the challenging NLP tasks such as question answering, machine translation, abstractive summarization, etc. have exponentially improved. However, with the state-of-the-art models, it is observed that even though these models generate natural and fluent-looking text but many times they are unfaithful and may contain facts/information that is irrelevant or not supported by the input. This phenomenon is referred to in the literature as a hallucination. A similar phenomenon is observed in end-to-end speech recognition systems, where the portion of the output text is having different acoustics when compared to the corresponding speech signal. In this tutorial, we introduce the problem of hallucinations in various Speech and NLP tasks such as machine translation, summarization and speech recognition. We categorize the hallucinations observed in this model and describe the techniques to quantify them. Next, we describe recent techniques to overcome hallucinations for many of these tasks. We draw the attention of the AI community to the potential problems of hallucinations in NLP and speech.