A Part-of-Speech Enhanced Neural Conversation Model.

Modeling syntactic information of sentences is essential for neural response generation models to produce appropriate response sentences of high linguistic quality. However, no previous work in conversational responses generation using sequence-to-sequence (Seq2Seq) neural network models has reported to take the sentence syntactic information into account. In this paper, we present two part-of-speech (POS) enhanced models that incorporate the POS information into the Seq2Seq neural conversation model. When training these models, corresponding POS tag is attached to each word in the post and the response so that the word sequences and the POS tag sequences can be interrelated. By the time the word in a response is to be generated, it is constrained by the expected POS tag. The experimental results show that the POS-enhanced Seq2Seq models can generate more grammatically correct and appropriate responses in terms of both perplexity and BLEU measures when compared with the word Seq2Seq model.