Research And Design Of Human-Computer Conversation Model Based On Deep Learning

It is a significant work in the domain of NLP to enable the interaction between human and machine.With the development of artificial intelligence technology,especially the wide application of deep neural network,the construction of dialogue system based on open-domain has attracted more and more attention.The development of the Internet has brought a great amount of accumulation of data,which makes the data-driven dialogue model possible.Now,most of the dialogue tasks are still under the framework of genera-tive tasks that are generally from end to end.The encoder-decoder model is an important mechanism for solving sequence to sequence tasks.It is the mainstream practice to extend and modify the encoder-decoder model to apply it to open-domain dialog tasks.Based on the encoder-decoder model,this paper proposes a new modification and supplement to apply it to single round dialogue and multi-round dialogue tasks.(1)For the single round dialogue task,the general RNN-based encoder-decoder model can not generate diverse and accurate responses well,so this paper proposes an enhanced-attention mechanism with sequential position information.The sequence information pro-cessed by GRU structure has sequential position information and is also used to calculate the internal attention value,which forms an enhanced-attention,fully expresses the rich information association within the sequence,and promotes the generation of diverse re-sponses.In decoding,the calculation of enhanced-attention makes the prediction of cur-rent step can receive the previous information according to different weight,which im-proves the accuracy of prediction at the end of the dialog sequence when using the RNN structure.Then,on the basis of the structure of transformer,aiming at the lack of location information,this paper proposes to add a sequential information module to the model and then applies the it to the single round dialogue task.(2)For multi-round dialogue task,the lack of effective data and too large structure of multi-round dialog model may lead to bad training.Based on the hierarchical recurrent encoder-decoder model,an auto-coding pre-training method is proposed.In pre-training,the latent variable mechanism in VAE is used to make the model learn various semantic information of the much normal dialog data in advance.In formal training,although the latent variables increases the diversity of semantic space,which makes there are more random choices in decoding to generate diverse responses,but the only randomness also makes the responses have the problem of information mismatch.To solve this problem,the improved model adds specific background information to the latent variables,such as topic or purpose information,in order to add some local deterministic factors to the global randomness.These operations enable the model to fully obtain the internal semantic information of dialog,and generate more relevant content.