Research On Generation Technology Of Domain-Specific Multi-Turn Dialogue

With the rapid development of deep learning technology and the wide application of artificial intelligence technology,multi-turn dialogue,which is an important task in natural language processing,has attracted the attention of researchers.Among them,domain-specific multi-turn dialogue generation has always been the focus of research in multi-turn dialogue.At present,the research of multi-turn dialogue generation in domain-specific has made remarkable progress and brings great convenience to people,for example,the application of voice navigation in the field of transportation,and the application of intelligent investment advisory in the field of finance,etc.However,there are still many challenges in the generation of domain-specific multi-turn dialogue.For example,the generated sentence is inconsistent with the dialogue scene;the construction of domain knowledge is difficult and the adaptability of domain is poor;poor memory of long-distance historical information;difficulty in generating factual long dialogue,etc.Therefore,the in-depth research on generation of domain-specific multi-turn dialogue has great research significance,not only for the promotion and development of natural language processing technology,but also for the landing and application of artificial intelligence technology in various fields.This paper mainly studied the generation technology of domain-specific multi-turn dialogue,and proposed corresponding solutions aimed at the problems existing in the generation of domain-specific multi-turn dialogue.Specifically,the main research contents of this paper can be divided into the following five aspects.(1)Firstly,in view of the shortage of resources in the judicial field,this paper published a large-scale dataset and defined several sub-tasks,such as fact-finding,role recognition,element recognition,feature recognition,intelligent dialogue,etc.,and provided baseline models for each task which can be further explored by researchers.In particular,for intelligent dialogue,this study further explored its applicability by using the transcripts of real world courtroom scenarios as training data.And investigated an intelligent body that automatically generated judges’ discourse through the dialogue between the judge,the plaintiff and the defendant.The agent can continuously ask questions to the plaintiff and the defendant based on their answers,so as to further explore the facts.Experiments on real world court transcript data demonstrated the superiority of the proposed scheme over traditional models.(2)The existing generative models usually adopted a sequence-to-sequence structure.However,the traditional sequence-to-sequence structure cannot identify the current scene information in a specific domain.Taking the judicial field as an example,according to the characteristics of generated discourse in different scenarios,this paper proposed a path-optimization knowledge fusion method.To be specific,the 283 judicial elements summarized by experienced judges were summarized in this paper as a logical tree,and each branch of the tree represented the key points of questioning in a case.In this paper,each element node on the tree was mapped to each sentence in multi-turn of dialogue,and then multiple nodes corresponding to dialogue were used as a knowledge path.Through path optimization,the optimized path was combined with the dialogue context as a kind of guidance information,so that the dialogue had a purpose,to determine what questions judges should ask in different trial scenarios.The experiment was carried out on the trial record data with more than 3 million sentences.It was proved that the proposed method was instructive and superior to the traditional model through two evaluation methods: manual evaluation and machine evaluation.(3)In the dialogue generation in a specific domain,the construction of external knowledge often costs a lot of manpower and material resources,and the domain adaptability was poor.Therefore,this paper proposed a cross copy network model and made similar dialogue as a kind of knowledge,the model can be copied from the context to generate the required entities,learned from similar dialogue inner logic,and can copy process framework,the framework can form the required sentence with entities copied from the context.Experiments on two datasets of justice and the e-commerce field proved that the proposed scheme can copy the important information of similar conversations and achieve the optimal effect through two evaluation methods: human evaluation and automatic evaluation.(4)Due to there were multiple turns of dialogue,the traditional model had a poor memory for long distance history information.Therefore,this paper proposed an deep reading memory network model,which used the last sentence of the current conversation context as Query,iteratively memorizing each sentence in similar conversations(as Key and Value),and finally integrating the memory information with the context to generate the current utterance.Experiments on two datasets of justices and the e-commerce field proved that the proposed scheme can memorize long distance history information and achieve the optimal effect through two evaluation methods: human evaluation and automatic evaluation.(5)The traditional pointer generation network can only copy a single entity,but it is also very important to copy phrases and sentences in domain-specific multi-turn dialogue generation.This paper proposed a multi-granularity copy network model,which can copy key entities from context and key phrase,sentences from similar dialogue.The model accomplished copy at three levels of granularity:words,phrases,and sentences,and the copy information was used to help generate the current dialogue.In this study,experiments were conducted on two datasets in the judicial and e-commerce domains,and the proposed method was demonstrated to copy important information from similar conversations with optimal results by both manual and automatic machine evaluations.