Research On Knowledge-driven Dialog Generation Model

In recent years,with the development of the Internet and social networks,a large amount of real social data has been accumulated,which effectively promotes the research of open dialog generation models.However,traditional dialog generation models assume that without additional background knowledge,it is easy to generate safe and meaningless responses,which cannot provide effective information to users.Therefore,numerous scholars and researchers have attempted to introduce external knowledge into dialog generation models to improve the accuracy and interest of response information,and extract and utilize external knowledge to improve the ability of dialog systems.Knowledge driven dialogue systems need to accurately convey information to meet user needs.Therefore,it imposes stricter requirements on the selection and use of knowledge.First,the system must select the most accurate knowledge fragment from a large external database,not just the relevant knowledge fragment.For example,in order to reflect the opening hours of a museum to users,reference materials from other museums should not be retrieved.Secondly,in order to generate accurate and coherent responses,the system needs to reasonably reason about the retrieved knowledge fragments and conversation context.In summary,this article proposes the following targeted improvement plans for the two issues mentioned above:1.Research on knowledge selection based on conversation context.Find the topics and issues involved in the conversation context,and query the corresponding knowledge in the triplet knowledge base based on the topics and issues.Due to an ASR error in the text,such as identifying "Paul Greengrass" as "Paul Greengrass "This article proposes a global pointer network model to effectively select topics in a conversation,and a multitask joint training model to select topics and issues contained in the conversation,which solves the potential mismatch between information used for topic reasoning and problem reasoning.Comparative indicators verify the effectiveness of the method proposed in this article in extracting conversation related knowledge.".2.Research on Knowledge Driven Dialogue Generation Strategies.Using external knowledge resources,retrieve relevant knowledge information and encode it into a model,and then generate responses to produce more meaningful results.However,these models cannot focus too much on external knowledge and ignore the conversation response of the other party.It is necessary to study appropriate knowledge fusion strategies.Therefore,we analyze the relationship between the generated conversation and knowledge information,predict the proportion of knowledge information required for the response message to be generated,and dynamically introduce knowledge information into the decoder state to predict the response.The experiment shows that the model achieves the best score on the Chinese conversation dataset,and the ablation experiment proves the effectiveness of the design of each module of the model.