Research And Implementation On Persona Based Personalized Dialog System

Dialog system is an important research direction in artificial intelligence field.In recent years,the development of Internet and Human-Machine conversation technology has grown to more application scenarios for dialog systems,and researchers have found that using persona information to improve dialog models increases the consistency and quality of the generated dialogues.Thus,fusing persona into dialog models and constructing more personalized and sophisticated dialog systems come with great research and application values.The thesis is mainly focused on two specific tasks for personalized dialog systems,which are personal attribute extraction task,and personalized response generation task.For the personal attribute extraction task,traditional sequence tagging-based information extraction models do not suitable for entity omitted dialog sentences since daily conversations are generally less rigorous,while current generation-based attribute extraction models have a lack in accuracy.The thesis improved a generation-based model which used a two-stage model for attribute extraction task,and integrated with a prompt revisiting module to make the model pay more attention on whole entity phrases rather than each single word.The experiments on two public data sets show that our model improved the overall accuracy in attribute extraction tasks compared with the baseline models,especially for unclearly expressed and entity omitted sentences.For the personalized response generation task,current models mostly depend on the selection of relevant personal attributes.Due to the low accuracy of persona selection,the quality in dialogue generation is also limited.The thesis improves a dialogue model based on Transformer structure,which uses all the candidate attributes as the input,making the model independent of the selection of relevant personalities.Furthermore,the model used a personality matching and revising module to make the model pay more attention on crucial personal attributes.The experiments on Persona-Chat dataset shows that the model increased on both conversation quality and personality scores compared with baselines.Finally,the thesis synthesizes both parts of the research.Firstly,the models and the data are both adjusted to Chinese conversation scenario.Then,the thesis design and construct a personalized dialog system.Finally,the models are tested and displayed systematically in Chinese conversation scenario.