Research On Dialogue Generation Task Based On Knowledge Graph

As an important branch of natural language processing,dialogue system has developed rapidly in recent years.Traditional dialogue systems are divided into task-based dialogue systems and open-domain dialogue systems.The former is mainly used to talk to and complete tasks with humans on specific tasks,such as booking assistants,intelligent customer service systems,etc.;the latter is mainly used to chat with humans.To allow dialogue systems to generate more meaningful responses,researchers try to incorporate external knowledge into the model,such as unstructured text and structured knowledge graphs.This thesis mainly studies the task of dialogue generation based on knowledge graph.Due to the differences in the dialogue history and knowledge graph structure,the existing methods mostly use the method of joint modeling through semantic latent vectors after encoding separately,which will not make full use of the topological structure of the graph for prediction;at the same time,the large-scale pre-training model has Developments have facilitated the application of pre-training plus fine-tuning models.However,when the application data of the model is limited to a specific domain,the vectors generated by the pre-training model may be aggregated in a highdimensional semantic space,which makes the model unable to make full use of the pre-trained information to improve the generation effect.In response to the above problems,this thesis aims to improve the accuracy and naturalness of the generated language to carry out algorithm research.The innovations are as follows:1)A dialogue generation algorithm based on knowledge-aware path is proposed.Through data analysis,it is found that entities and relationships are the logical intersection of multiple rounds of dialogues and knowledge graphs,and the round transfer of conversations also corresponds to the transfer of entities on the knowledge graph.Therefore,this thesis attempts to build a knowledge-aware path to jointly model the dialogue history and knowledge graph.By identifying the entities and relationships mentioned in each sentence,the conversation round information is constructed in the knowledge graph to build a migration path.Finally,from the perspective of graph topology Make predictions about knowledge choices to assist dialogue generation.2)A dialogue generation algorithm for structured path modeling is proposed.According to the different processing methods of knowledge-aware path,this thesis proposes a dialogue generation algorithm for structured path modeling.This method can obtain more fine-grained knowledgeaware path information from the dialogue history.Experimental results show that the proposed method outperforms the baseline model in both knowledge selection effect and dialogue generation accuracy.3)A dialogue generation method based on contrastive learning is proposed.This method draws on the ideas of text data enhancement and recall negative sampling of recommendation system,and constructs a knowledge-related unsupervised dialogue positive and negative sample enhancement algorithm.The contrastive learning pre-training module is introduced on the input side of the existing model,and the encoder parameters are pre-trained based on contrastive learning.The experimental results show that the method proposed in this thesis alleviates the clustering of sentence vectors in the semantic space,and the model can generate more diverse responses.