One-Versus-AllとAttention機構を取り入れたRNNによる対話行為推定

抄録

For understanding the contents of a user's utterances, dialogue acts classification is often introduced to chat systems. This paper proposes Enhanced One-versus-All RNN (ENOVA RNN), which is a dialogue act classification model that consists of two RNN layers, one-versus-all layers, and the attention mechanism. The content of past utterances in dialogue is an important feature for dialogue acts classification models. In addition to the fact, classifiers tend to confuse dialogue acts that rarely appear in dialogue with frequent ones because the number of utterances differs greatly in each dialogue act. ENOVA RNN is a classification model to capture some features of dialogue acts that rarely appear after considering the content of past utterances in dialogues. In this study, it was confirmed that ENOVA RNN can classify dialogue acts using contexts not greater than six sentences. Moreover, ENOVA RNN improves rare dialogue acts classification performance keeping the overall quality of the performance by narrow down the dialogue acts using attention weights.