Investigating and Addressing Hallucinations of LLMs in Duties Involving Negation
Authors: Neeraj Varshney, Satyam Raj, Venkatesh Mishra, Agneet Chatterjee, Ritika Sarkar, Amir Saeidi, Chitta Baral
Summary: Massive Language Fashions (LLMs) have achieved exceptional efficiency throughout all kinds of pure language duties. Nonetheless, they’ve been proven to endure from a crucial limitation pertinent to ‘hallucination’ of their output. Latest analysis has centered on investigating and addressing this drawback for quite a lot of duties resembling biography era, query answering, abstractive summarization, and dialogue era. Nonetheless, the essential side pertaining to ‘negation’ has remained significantly underexplored. Negation is essential as a result of it provides depth and nuance to the understanding of language and can be essential for logical reasoning and inference. On this work, we tackle the above limitation and notably concentrate on learning the influence of negation in LLM hallucinations. Particularly, we examine 4 duties with negation: ‘false premise completion’, ‘constrained reality era’, ‘a number of selection query answering’, and ‘reality era’. We present that open-source state-of-the-art LLMs resembling LLaMA-2-chat, Vicuna, and Orca-2 hallucinate significantly on all these duties involving negation which underlines a crucial shortcoming of those fashions. Addressing this drawback, we additional examine quite a few methods to mitigate these hallucinations and show their influence.
