Neural Conversation Models and How to Rein Them in: A Survey of Failures and Fixes

TL;DR

This survey reviews neural conversation models, analyzing their common failures and proposing fixes by examining intervention strategies at data, training, and decoding stages to improve conversational quality.

Contribution

It systematically categorizes recent approaches to improve neural dialogue systems by addressing issues of fluency, informativeness, and social appropriateness, and suggests directions for future research.

Findings

Neural models often produce fluent but inconsistent responses.

Interventions at data, training, and decoding stages can improve conversation quality.

Future research should focus on integrating social norms and coherence.

Abstract

Recent conditional language models are able to continue any kind of text source in an often seemingly fluent way. This fact encouraged research in the area of open-domain conversational systems that are based on powerful language models and aim to imitate an interlocutor by generating appropriate contributions to a written dialogue. From a linguistic perspective, however, the complexity of contributing to a conversation is high. In this survey, we interpret Grice's maxims of cooperative conversation from the perspective of this specific research area and systematize the literature under the aspect of what makes a contribution appropriate: A neural conversation model has to be fluent, informative, consistent, coherent, and follow social norms. In order to ensure these qualities, recent approaches try to tame the underlying language models at various intervention points, such as data, training regime or decoding. Sorted by these categories and intervention points, we discuss promising attempts and suggest novel ways for future research.