Topic Expansion Method Considering Randomness for Dialogue System
Yoshitaka Mikami, Masafumi Hagiwara · Transactions of Japan Society of Kansei Engineering · 2018
In a non-task oriented dialogue system, it is important that the system can expand various topics for continuing a conversation. At the conventional dialogue systems, topic expansion has not been paid attention. The conventional dialogue systems often use the user's input topic word as a topic word of system's utterance. In this study, we propose a topic word expansion method. Concretely, we extract a topic word from user's utterance and generate embeddings using genetic algorithm and embeddings of the topic word from the user's utterance. As the evaluation function of genetic algorithm, we use neural network that learns topic word correspondence in a dialogue corpus, lexical knowledge and heuristic rules. By using the proposed system, we aim to model ambiguity of topic words and make it possible to expand topic words of conversation integrating plural knowledge. As evaluation experiments, we evaluated the proposed system by subjectivity and verified the effectiveness.