Bilingual Japanese/English robot dialogues with knowledge graphs and conversational AI

Abstract

The paper describes recent work on robot dialogue systems that use knowledge graphs as their primary source of information. The aim is to make generic dialogue systems by putting domain knowledge in knowledge graphs and adding ontological metadata such as taxonomies. A demonstration prototype uses the classic restaurant search domain. The robots answer user queries by searching knowledge graphs stored in graph databases, with multilingual labels for restaurants, cuisines and other objects. The system uses open source transformer-based conversational AI to train models for natural language understanding and for dialogue response selection. The paper describes new work on a bilingual version of the system using an experimental mixed-language NLU model trained with both Japanese and English examples. With the mixed-language model the system can recognise user intents and mentioned entities from queries in either Japanese or English. By default the robots give bilingual responses in both Japanese and English, switching to monolingual responses on request. To avoid sudden changes of voice, the robot uses the same Japanese voice to speak Japanese and English.