Open-domain Utterance Generation for Conversational Dialogue Systems using Web-scale Dependency Structures

Even though open-domain conversational dialogue systems are required in many fields, their development is complicated because of the flexibility and variety of user utterances. To address this flexibility, previous research on conversational dialogue systems has selected system utterances from web articles based on surface cohesion and shallow semantic coherence; however, the generated utterances sometimes contain irrelevant sentences with respect to the input user utterance. We propose a template-based approach that fills templates with the most salient words in a user utterance and with related words that are extracted using web-scale dependency structures gathered from Twitter. Our open-domain conversational dialogue system outperforms retrieval-based conventional systems in chat experiments.