Language-independent Dialogue Data Filtering for Neural Dialogue Response Generation

Abstract

In the area of sentence generation using deep neural network technology, e.g., machine translation, automatic summarization, and dialog response generation, approaches to increase the performance of models by improving the quality of training data have been spotlighted. In this paper, we propose a scoring function that detects low-quality utterance-response pairs in training data to improve the performance of a neural dialogue response generation model. Specifically, our function combines two viewpoints, "typical phrase interconnection" and "topic consistency", to rate the plausibility of two consecutive utterances as dialogue. In our experiments, we apply the proposed method to conversation data in multiple languages and demonstrate that the proposed score is correlated with human subjective ratings. Moreover, we demonstrate that training data filtering with our score is effective for improving the performance of response generation models using automatic evaluation and manual evaluation.