Resolving Context Contradictions in the Neural Dialogue System based on Sentiment Information

Abstract

Chatbots trained on large corpus generate fluent responses, but often suffer from the problem of generating responses that contradict past utterances. Recent research treats dialogue contradiction detection as a task of natural language inference (NLI), and a method to remove contradiction from responses has been proposed and has shown high performance. However, these datasets do not provide explicit information about emotions, and these models cannot capture changes in emotions. In this work, we create a new dataset by explicitly labeling emotional information on an existing contradiction detection dataset and use this dataset to train an NLI model. Furthermore, we train the NLI model on the original dataset as well and compare the accuracy of both in dialogue contradiction detection.