Dialogue act estimation using deep learning based parser in negotiation dialogue system

Abstract

In recent years, research on dialogue agents has been actively conducted. One of the previous studies is a method that decouples strategy and generation using dialogue act that represents the intention of the utterance. This method using dialog acts in the sentence generation process improved the task success rate and the human-likeness of the utterances. On the other hand, since the parser is implemented rule-based, there is a limitation to the sentences it can parse the dialogue acts. Based on the above background, in this study we annotate training data based on the proposed dialogue acts and propose a parser based on deep learning. The parser using deep learning shows that the accuracy of classification for dialogue act is approximately 83%. Furthermore, we succeeded in reducing the number of unknown dialogue act.