Recognising flexible intents and multiple domains in extended human-robot dialogues

抄録

In earlier work we developed two dialogue systems for Nao robots. WikiTalk (COLING 2012, COLING 2016) supports Wikipedia-based dialogues on open-domain topics. CityTalk (IJCAI 2018) supports task-based dialogues on hotel and restaurant domains, and can change domains during the dialogue. We focus on extended multi-turn multi-topic human-robot dialogues, which are more challenging than short question-answering interactions. We used WikiTalk with ERICA at Kyoto University (IJCAI 2019), and aim to make our systems available across multiple robot platforms. To support this objective we are using Rasa, an open-source conversational AI framework, with encouraging results. Rasa supports transformer-based dialogue models which can recognise flexible intents in extended multi-turn multi-domain dialogues, not merely pre-specified intents within a single domain. We also aim to extend CityTalk to more than two domains. To support this objective we are using MultiWOZ, a large-scale multi-domain dialogue dataset. Later we aim to combine our systems in a single system that can handle task-oriented dialogues like restaurant searches and chat-oriented dialogues using Wikipedia information.