Roving Mind: a balancing act between open–domain and engaging dialogue systems

Open–domain dialogue systems should be able to cover a very large set of domains and at the same time keep the user engaged in the interaction. Current approaches to dialogue modeling are divided between domain–independent, non–modular approaches using sequence–to–sequence models and the domain–specific modular systems developed for task–based dialogue. Furthermore, user engagement in dialogue, addressed in this bot challenge, is a rather new research area. In this work we describe Roving Mind, a dialogue system which combines domain– independence with a modular architecture for open–domain spoken conversation, with a specific module for user engagement. Our architecture takes a balanced approach between human expert design and data–driven approaches, adapting the traditional task–based architecture relying on slots and intents to open–domain conversation using entities and domain-independent dialogue acts. The system was tested extensively during the Alexa Prize semifinals and received ratings from a large number of users, which allowed us to draw statistically relevant facts on how users interact with an open domain, non task-based social bot. Our experiments show that on average users give higher ratings to task-driven conversations (using user–entertaining strategies) compared to completely open–topic conversations. We also find a correlation between cumulative sentiment (using sentiment and dialogue acts) and user ratings and argue that this could be investigated to estimate an error–signal for strategy computation.