Human-computer dialogue breakdown detection

Abstract

With the steady growth in the use of consumer relationship technologies on the Internet, chatbot systems have become ubiquitous in Natural Language processing (NLP) and related fields. Despite significant advances in recent years, however, systems of this kind do not always deliver plausible, consistent results, in many cases leading to a dialogue breakdown. As a result, there is a growing interest in how to improve systems of this kind so as to minimise errors. Based on these observations, this work addresses the issue of automatic dialogue breakdown detection by presenting three models that take the dialogue history into account to decide when a conversation is likely to break. The models under consideration explore a range of recent NLP methods and are evaluated by using a purpose-built Portuguese dataset conveying real-world human-computer conversations, and also in publicly available benchmarks for the English language.