Chat Response Generation Based on Semantic Prediction Using Distributed Representations of Words

In this paper, we propose a chat response generation system using distributed expression of words by Word2vec. With the conventional one-hot representation method, there was a problem in that the model becomes complicated as the vocabulary increases, and only the words that appear in the dialogue corpus can be handled. We address these problems by using Word2vec and extend it to handle unknown words that did not appear in the conversation corpus. In a subjective evaluation experiment, we showed that various responses can be generated by estimating words using semantic prediction.