TIER: A Novel Model Based on POS Information to Generate Dialogue

The design of dialogue system which interacts with users in natural language ranks high on the agenda of current NLP research. A well-performed dialogue system should generate responses that are not only diverse and meaningful but also grammatically correct. Although the former gives a rise to a research hotspot, the latter hasn’t gotten as much attention. In this paper, we proposed POS-aided model called tag guided encoder-decoder, which takes not only word sequence but also POS tag sequence as inputs. By introducing POS tags, the model can operate on part-of-speech information implicitly, and thus gains the syntactic guidance to generate grammatical responses. In our experiment, we apply the model to dialogue response generation in the open-domain and compare it with competing models. The results show that our model outperforms the competing model according to BLEU metric and human evaluation study.