Towards Social and Interpretable Neural Dialog Systems

Abstract

Open-domain dialog generation is a task that challenges machines to mimic human conversations. Despite the remarkable progress natural language generation has seen over the past several years, open-domain dialog systems still suffer from limitations that hinder their adoption in the real world. Systems trained with maximum likelihood often generate dull and repetitive responses, ignoring user input. Training on standard datasets from online forums leads to the generation of inappropriate, biased, or toxic responses. And models rarely exhibit long-term coherence across multiple dialog turns. Meanwhile, the predominant approach to dialog generation relies on black-box neural networks which provide little insight as to what information they learn (or do not learn) about engaging in dialog. In light of these issues, this thesis makes two contributions to building social and interpretable dialog systems. The first part of this thesis proposes a novel reinforcement learning approach for improving the social capabilities of open-domain dialog systems. We optimize for human-centered objectives such as response politeness, diversity, coherence, and sentiment. Our interactive human evaluation shows that these objectives can improve the quality of human-AI interaction and increase user engagement. The second part of this thesis investigates the conversational understanding captured by neural dialog systems using probing. Our results suggest that standard open-domain dialog systems struggle with basic skills such as answering questions, inferring contradiction, and determining the topic of conversation. We also find that the dyadic, turn-taking nature of dialog is not fully leveraged by these models. By exploring these limitations, we highlight the need for additional research into architectures and training methods that can allow for capturing high-level information about natural language.