Research On Topic Guidance Of Open Domain Dialogue Generation

With the development of social networks,various social platforms have gradually penetrated into people's lives.In addition to offline communication,people's social way has gradually shifted to online communication.This transformation has accumulated a large number of real dialogue resources in the social platform,which not only provides a large number of corpus for the research of open domain generative dialogue,but also makes the construction of automatic dialogue system feasible.More importantly,the dialogue rules and models contained in real dialogue resources will promote the research of automatic dialogue technology and promote the solution of key problems in the dialogue system.In addition,there are more and more application scenarios of open domain generative dialogue,such as automatic customer service system and personal assistant,which can save resources and improve efficiency for these scenarios.Therefore,a profound study of open domain generative dialogue plays a vital role in both the research of natural language processing technology and the development of Internet related industries.This paper focuses on the topic guidance method of open domain dialogue generation.Aiming at the problem that the current research on dialogue generation ignores the importance of topic to dialogue generation,this paper proposes topic guidance methods for single-round and multi-round of dialogue respectively.Specifically,the main research content of this paper has the following two aspects.(1)Single-round dialogue generation model based on dual confrontation network.The topic of the dialogue content reflects the semantic logic of the dialogue process and is the key information of the dialogue data.The dialogue generation model based on seq2 seq can effectively obtain the sequence relationship between words,but ignores the topic logic between sentences.The generated response is prone to meaningless sentences(such as "I don't know"),which is called "safe reply" problem.In this paper,a single-round dialogue generation model(TR-ARG)of dual confrontation network is proposed.The topic attention mechanism is embedded in the generation model to guide the model to produce a response in line with the topic logic.The model design adopts the dual confrontation network mode to jointly optimize the text sequence and topic logic objectives.A comparison index is designed on an open Opensubtitles conversation dataset to verify the effectiveness of the proposed method for topic-oriented tasks in a single round of conversations.(2)Research on multi-round dialogue generation strategy based on topic guidance.Natural dialog is a multi-round dialogue process produced by many alternating speeches.With the evolution of the speech process,the content of the conversation and the topic migrate synchronously.How to control the topic trend of the generated responses is a problem that must be considered in the dialogue generation.The existing multi-turn conversation model still lacks effective chat topic control mechanism,which makes the model lack topic inference ability and easy to generate monotonous and discontinuous sentences.This paper presents a topic-guided multiple conversation generation strategy(DATHRED).A topic smoothing strategy based on knowledge map is designed in the generation model to improve the topic inference ability of the model.Based on the dual antagonism,the model objective function is designed to complete the model parameter learning.For the proposed topic smoothing strategy,a topic smoothing evaluation method is designed experimentally,and a comparative experiment is designed on the Kd Conv dataset,in order to verify the effectiveness of the proposed method for topic-oriented tasks in multi-round conversations.The method proposed in this paper makes a new exploration on the topic smooth reply method of open domain dialogue generation task at the theoretical level,and provides a new technology and model for the application of man-machine dialogue system in e-commerce at the application level,which has good theoretical significance and commercial promotion value.