Graph Functional Dependency-based For Consistent Dialogue Generation

The consistency of dialogue system means that the semantics of the current response in the dialogue system is consistent with that of the historical responses.That is,there is no conflict.In the consistency dialogue,the logical conflict of attribute semantics is one of the main reasons hindering the generation of the consistency dialogue.The previous works used the encoder-decoder structure to learn the semantics of attributes implicitly.However,this method cannot accurately describe the semantics of attributes to maintain consistency.The subsequent works are dedicated to limiting the semantics to the artificially defined precise space to generate similar attributes to avoid conflicts.However,this method has poor generalization,especially when extended to open fields,it cannot effectively set the predefined semantic space,and contradictions are unavoidable.The scientific problem of this paper is how to accurately represent the semantics of attributes in a semantic space that is not artificially predefined and to ensure the consistency of the semantics of attributes of the generated responses.Graph Functional Dependencies(GFDs)have strong advantages in consistency recognition and extraction.However,it is a big challenge to use dialogue data to learn the semantic of attribute relationship between graph nodes.To solve the above challenges,this paper proposes the GFDC-Att model,which is divided into two aspects,focusing on consistency and fluency,respectively,and improving the generalization ability of the dialogue system when maintaining consistency.First of all,for consistency,this paper proposes a method based on Graph Functional Dependencies to guide the generation of consistent responses.To extract features accurately,use the Commonsense Knowledge Base to abstract the graph data constructed by historical responses,making attribute semantic conflicts more prominent.In addition,the consistency feature information provided to the generator by the consistency module based on the Graph Functional Dependencies is merged to generate a consistent response.In addition,since the introduction of the consistency feature affects the encoderdecoder information of the dialogue system,it is not conducive to generating smooth responses.To solve this problem,this paper also proposes a Sequence Generation Adversarial Network(Att GAN)based on the attention mechanism in the GFDC-Att model.The attention mechanism is introduced to solve the Monte Carlo search’s high computational cost and sampling error.The attention mechanism is introduced to solve the problems of high computational cost and sampling error of Monte Carlo search in the original model.An attention layer is added to the discriminator,and the attention score is used as the basic reward so that the discriminator can calculate the reward value of each word.And to meet the requirement of effective reward value,the attention score is scaled,and the appropriate reward value is obtained through function mapping to improve the fluency of generating responses.Finally,this paper designs a large number of comparative experiments on the proposed model.The experimental results show that the consistent dialogue generation method has a good generation effect,especially in maintaining the consistency of the responses.The model has achieved excellent performance.This paper also conducts a large number of experiments on Att GAN.The experimental results show that Att GAN can speed up training and improve the quality of generation.Finally,the Att GAN is combined with the consistent dialogue generation model to improve fluency.The GFDC-Att model is experimentally verified,and the experiment shows that the response quality is further improved.