Automatic Answering of Dialogue Completion Problems Based on Naturalness and Fluency of Sentences

Abstract

In the “Can a Robot get into the University of Tokyo?” project, we tackle the dialogue completion problem in English. Fine-tuning was performed for XLNet, RoBERTa, or ALBERT using automatically generated psudo- problems. The model determines whether the conversation sentence is natural and selects an appropriate word from the options in the blank. It was found that the correct answer rate was improved by more than 10 points in all models by using the speaker information rather than inputting only the conversation sentence. In particular, the system using ALBERT achieved a very correct answer rate of 0.75.