Unsupervised Question Answering: Challenges, Trends, and Outlook

Abstract

Question answering (QA) is considered to be a central aspect of natural language processing (NLP) and has seen remarkable progress in the last decade, brought-about by transformer-based language models trained on large human-annotated text corpora. However, several pitfalls of supervised training have been identified, especially when considering performance of such systems on new domains, linguistic styles, and adversarial samples. Unsupervised question answering – the ability to answer questions without explicit supervision from human-annotated training data, has emerged as a research direcftion that could potentially mitigate these pitfalls. This paper reviews recent trends in unsupervised question answering and provides a unifying perspective of work in this area, along with a survey of the closely related directions of weakly and partially supervised QA models. We provide insights into associated challenges and potential research directions towards robust unsupervised QA models.