[3O1-GS-13] AI application: Dialogue system and mind

[3O1-GS-13-01] Development of Dialogue System for Controlling End of Sentences based on Politeness Theory

〇Nozomu Nagai¹, Tomoki Miyamoto¹, Daisuke Katagami¹ (1. Graduate School of Tokyo Polytechnic University)

Keywords:Sequence to sequence, Non-task-oriented dialogue system, Politeness theory

In this study, based on the politeness theory which is known as linguistic consideration to construct smooth human relation by the speech and behavior of each other, the purpose is to develop 2: "Politeness word priority system" which preferentially outputs negative politeness (NP) words such as positive politeness (PP) words and honorific words used in friend tone, etc., and "Politeness utterance learning system" which classifies sentences based on PP and NP words and learns them for each classification. The politeness word priority system extracts PP and NP words and preferentially speaks them. The politeness utterance learning system classifies sentences of data sets according to whether PP and NP words are used in sentences, and learns them for each classification. The developed system uses about 1.15 million pairs of data of tweets and replies on Twitter, and learns appropriate replies to utterances by a sequence to sequence model. As a result of an impression evaluation experiment on whether a politeness word priority system and a politeness utterance learning system could carry out an appropriate answer, it was proven that a politeness word priority system gave an impression of pessimism and empathy, and a politeness utterance learning system gave an impression of familiarity and politeness.