Transfer Learning from Other Languages to Japanese Dialogue Response Generation

Abstract

Compared to English and Chinese, there are not many high-quality publicly available corpora for Japanese in terms of chat dialog response generation. Therefore, in order to achieve a high enough performance in chat dialogue generation with small and low-quality data, this study utilized transfer learning from Chinese and English in the Transformer-based model. Three Japanese corpora and a corpus collected from Twitter were used as the dataset to generate the input sentences. The average value of the distinct-1 as an automatic evaluation index of the generated results was 0.368 without transfer learning, and 0.412 for the transfer learning model. In terms of human evaluation, the model with transfer learning scored significantly better on all three items: sentence connection, informativeness, and humanness, compared to the model without transfer learning, for a small training data set with 9343 sentences of dialogue.