Abstract
In this paper, we propose implicit and explicit utterance generation models and a dialogue system in which such methods are implemented. Modularizing classifiers enables the agent to give input utterance tags of multiple features including types of sentences and mood expressions. The explicit responses are generated if the input text is classified in a specific domain such as Question-Answering, based on the tags given by classifier modules. In the implicit way, the features gotten from the inputted sentences define an agent's internal state. A relativity vector to each domain is sustainably computed based on similarity in Japanese WordNet ontology as the system's internal state. In other cases of the classification, the system generates open-domain utterances. We will discuss the result of experiments intended to show characteristics of both domain detection methods.
