Research On Single-round Dialog Generation Based On Recurrent Neural Network

Human computer conversation is a frontier research topic in the field of artificial intelligence,which involves language understanding,reasoning and use of common knowledge.With the development of deep learning technology,the scientific medias and the researcher communities keep an eye on dialogue.And many excellent systems use recurrent neural network(RNN).It is of great research significance to apply deep learning to dialogue.In this thesis,we consider a much-simplified version of the open domain dialogue system which is Short Text Conversation(STC).STC means one round of conversation formed by two short texts that does not limit the chat scene,which is the core problem in the research dialogue system.We solved the STC based on RNN and the approaches can be divided into two categories: retrieval-based STC and generation-based STC.Given a query,the retrieval-based system retrieves candidate responses from the large repository of conversation data.And then the system compares their text features by text matching models and returns the most reasonable response.Experiments have compared various text matching models,including keyword matching,shallow semantic matching and deep semantic matching model based on RNN.Experimental results show that semantic matching has significantly improved the matching effect in dialogue system.In this thesis,we also adopt an encoder-decoder RNN-based structure to build and train a Chinese generation-based STC.Experiment shows that the generative models are able to generate proper responses that could have never appeared in the corpus.But there are many generalizing responses in the generated sentences.To solve the problem of how to generate more quality responses requires more research and exploration.