2016 年 28 巻 1 号 p. 535-555
This article presents an innovative dialogue agent designed for textual casual chatting, which can handle any language. The system acquires knowledge from a non-annotated corpus and then represents all the language aspects as a graph. Using graph traversal, the system generates one or several outputs corresponding to the user's input. Moreover, it uses graph clustering to generate word categories without using any grammar information, and finally uses these to generate more various responses. In addition, all the operations are processed in parallel, making the system able to process any input in real-time, as in human conversations. Since the system accepts any kind of input, it can also be considered to be naturally multimodal. We carried out experiments in Chinese, English, French, Japanese and Korean, and obtained results comparable to a more language-specific multilingual system. Although the current system deals only with a limited corpus and as a consequence only handles simple dialogues, we demonstrate that with further interaction and language samples it is able to adapt to more sophisticated dialogues.
