The Research On Knowledge-driven Open Domain Dialogue Generation Methods

Open-domain dialogue system is one of the important tasks of interest in the fields of human-computer interaction,intelligent science,and natural language processing.It has received a wide range of attention from academia and industry for its excellent commercial value in areas such as chatbots,virtual personal assistants,and medical support.Open-domain conversations aim to allow machines to understand and use natural language to interact with users with the same thoughts of human communication,which requires that the generated utterances not only conform to the contextual semantics and basic grammatical rules but also reflect a sufficient variety of content.Language expression is inseparable from knowledge.Borrowing background knowledge for thinking in dialogues is an important feature of human communication intelligence.In recent years,researchers have devoted themselves to introducing external knowledge bases in open-domain dialogue systems to bridge the background knowledge gap between machines and humans,as a way to improve the performance of dialogue models.However,the performance of existing knowledge-driven dialogue models in complex scenarios is still unsatisfactory,and the existing work mainly suffers from the following two deficiencies: 1)The accuracy of knowledge selection is low.When facing complex dialogue scenarios such as dialogue subject shifting,the existing knowledge selection methods lack the reasoning ability and are difficult to generalize to the correct knowledge based on the current dialogue;2)The efficiency of knowledge fusion is insufficient,and the existing model lacks the dynamic attention to the internal details of knowledge,which makes it difficult to generate rare vocabulary in knowledge.In this paper,we will take both knowledge selection and knowledge fusion in opendomain dialogue systems as the entry point,to investigate the knowledge-driven opendomain dialogue methods.The main research content includes:1.We propose a divergent knowledge selection method with one-hop reasoning for improving the robustness and generalization of the knowledge selection in complex dialogue scenarios.To solve the knowledge selection boundary problem caused by subject shifting,we construct a knowledge optimization module based on topic association information,and also design a teaching framework and content constraints to guide the inference of this association information.Knowledge optimization is used to reduce the influence of subject shifting,and improve the accuracy and robustness of knowledge selection under complex dialogue scenarios.2.We propose a knowledge-aware dialogue generation method based on the attention mechanism.Compared with most current knowledge fusion methods,our method introduces a novel knowledge attention sublayer in the Transformer decoder.This attention sublayer models the correlation between knowledge and responses,aiming to dynamically focus on the internal details of knowledge as well as capture the key information in the knowledge.Then,in the process of response word prediction,our method also includes a copy-enabled dual-source pointer network,which directly embeds the rare words in knowledge into the response by word copying.It accelerates the model convergence speed and also enhances the model's ability to generate sparse words.We experimentally validate the effectiveness of the proposed method on the recently released Wikipedia knowledge grounded conversation dataset Wizard-of-Wikipedia.The experimental results show that our proposed method achieves leading performance in various evaluation metrics,including quantitative and manual,and can generate more natural and fluent response statements with more rich knowledge content.