Research On Dialogue Generation Methods Based On Neural Networks

Dialogue is the important ways of communication between people,and it is also the main means of human-machine interaction.With the development of artificial intelligence technology,dialogue generation technology has also developed rapidly,but it still can not meet people’s growing needs.The research on the new technology of dialogue generation based on neural network has theoretical value and practical guiding significance for promoting the development of dialogue system and human-machine interaction technology.So far,the dialogue generation method based on deep neural network has made great progress,but there are still some problems to be explored:(1)the existing end-to-end dialogue generation system is easy to generate generic responses,and the generated responses contain a small amount of useful information and poor diversity.Therefore,how to generate a diverse responses rich in more useful information has become an urgent problem in practical application.(2)The communication between people is rich in emotion.However,the existing dialogue generation models lack empathy.Therefore,how to construct the multi-turn emotional dialogue generation model has become a new research hotspot.(3)The dialogue generation model with superior performance depends on large-scale and high-quality dialogue corpus.Especially in the face of new fields,less dialogue corpus will lead to serious performance degradation.Therefore,the research of dialogue generation method in low resource scenario has become an urgent problem to be solved.Facing the existing problems and new challenges of dialogue generation,this paper deeply studies the dialogue generation method based on neural network,and focuses on the generation of dialogue generation model rich in useful information and diverse responses,multi round emotional dialogue model and low resource dialogue generation solutions.The research methods and results are summarized as follows:1.Aiming at the problem that the response generated by the existing dialogue generation model has less useful information and poor diversity,two topic-aware hierarchical latent variable dialogue models(VHCR-T)are proposed,namely,the topic-aware model with attention mechanism and the topic-aware model with dual topic latent variables.Both models perceive the topic information contained in the context by extracting features from the topic level information,and then input the extracted features into the decoder to generate a response containing more information.At the same time,the sentence level latent variables used in the two models can increase the diversity of responses.The experimental results show that compared with the baseline models,VHCR-T can effectively improve the content of informativeness and the diversity of responses.2.In order to further improve the content of useful information in the response and maintain the diversity of response,an adversarial dialogue network(SDAN)based on knowledge and diverse syntax is proposed.Considering the introduction of knowledge in the knowledge graph to increase the informativeness of response,but the semantic latent variables used to maintain the diversity of response may lead to the inaccuracy of knowledge decoding in the knowledge graph.Therefore,in order to achieve a balance between increasing diversity and maintaining the accuracy of knowledge decoding,The network introduces diverse syntactic information to generate syntactic diverse responses without affecting the accuracy of knowledge decoding.In addition,SDAN also introduces the confrontation generation network to the semantic encoding module to ensure that the semantic encoding module does not contain syntactic information,so as to maintain the controllability of syntax.Experiments show that SDAN can not only maintain the accuracy of knowledge decoding,but also improve the diversity of responses on the basis of maintaining the same response semantics.3.Aiming at the problem of generating responses with empathy,a multi-turn emotional conversation model(MECM)is proposed.Based on the hierarchical latent variable model,an emotional latent variable is added to model the emotional transfer process between contexts.At the same time,an emotion classifier is introduced,which is used to increase the emotion recognition ability of the model in the process of dialogue on the one hand,and to guide the latent variables of emotion on the other hand.In addition,the two tasks of emotion recognition and dialogue generation are trained at the same time to increase the effect of dialogue generation by utilizing multi-task learning.The experimental results show that compared with the baseline models,MECM can not only improve the semantic similarity of response,but also greatly improve the diversity of responses and the accuracy of emotional expression.4.Aiming at the problem of dialogue generation in low resource scenarios,a multi-source low resource dialogue generation model based on inverse curriculum learning is proposed.Firstly,the model utilizes paraphrase generation model,round-trip translation model and pre-training dialogue generation model to generate augmented data based on source data,and integrates the source data with augmented data for training.Then,three training strategies of curriculum learning,inverse curriculum learning and curriculum learning + inverse curriculum learning are designed to train the fused data.Finally,the source data is used to fine-tune the trained model.The experimental results show that compared with the baseline models,the proposed model has significantly improved in semantic consistency and diversity.