THINK: A novel conversation model for generating grammatically correct and coherent responses

Many existing conversation models that are based on the encoder–decoder framework incorporate complex encoders. These powerful encoders serve to enrich the context vectors, so that the generated responses are more diverse and informative. However, these approaches face two potential challenges. First, the high complexity of the encoder means relative simplicity of the decoder. There is a danger that the decoder becomes too simple to effectively capture previously generated information. As a result, the decoder may produce duplicated and self-contradicting responses. Second, by having a complex encoder, the model may generate incoherent responses because the complex context vectors may deviate from the true semantics of context. In this work, we propose a conversation model named “THINK” (Teamwork generation Hover around Impressive Noticeable Keywords) that is equipped with a complex decoder to avoid generating duplicated and self-contradicting responses. The model also simplifies the context vectors and increases the coherence of generated responses in a reasonable way. For this model, we propose Teamwork generation framework and Semantics extractor. Compared with other baselines, both automatic and human evaluation showed the advantages of our model.