Embeddings Based Dialogue System for Variety of Responses

Abstract

In this paper, we propose an embeddings based dialogue system for variety of responses. Recently, embeddings have been widely used to contribute to various kinds of natural language processing tasks. However, embeddings are not employed as internal expression in a dialogue system. In the proposed system, we newly employ template features as embeddings by using a neural network and combine them with the conventional methods that specialize in embeddings. By using embeddings as an internal expression, it is expected that it can acquire generic features from fewer learning data to construct a dialogue system using fewer rules. As an experiment, we carried out a subjective evaluation. As a result, the proposed dialogue system using a few learning data obtained evaluation value almost equivalent to the conventional state of the art dialogue systems.