| As artificial intelligence technology rapidly advances and society becomes increasingly intelligent,people have raised higher expectations for intelligent systems.As an important human-computer interaction mode,multi-turn dialogue systems have been a hot topic of concern for academia and industry both domestically and abroad,and are considered to be one of the important technological challenges facing human-computer intelligent interaction.Current research on multi-turn dialogue generation is mainly based on sequence-to-sequence dialogue generation models.However,these models tend to generate safe replies with little diversity and information,resulting in meaningless and boring interaction.This deficiency is especially acute when participants try to discuss in-depth topics on a particular domain.Therefore,more and more researchers have begun to study knowledgedriven dialogue system,exploring the ability to improve dialogue systems by introducing external knowledge.The two key issues of the task are:(1)How to select the specific knowledge related to the dialogue context from the huge external knowledge(2)How to reasonably incorporate selected knowledge to generate context-sensitive and highly informative responses.In order to solve the problem of insufficient information and diversity of responses generated by multi-turn dialogue systems,this topic introduces external knowledge graphs into dialogue systems,and conducts research on two key issues of knowledge-driven dialogue tasks.The main contributions of our thesis are as follows:1.A dialog context-aware knowledge selection model is proposed.Acquiring dialogue context-aware dialogue history codes for knowledge matching,and improving the accuracy of knowledge selection by judging whether external knowledge is needed and using replies as posterior supervisory signals during training.The experimental results,when compared with baseline models show that the proposed model can more accurately select the knowledge related to the dialogue according to the current dialogue context,and improve the quality of the overall reply generation.2.A knowledge-fused multi-turn dialogue generation method based on pre-trained dialogue model is proposed.By adjusting the input encoding of the pre-training model to distinguish the knowledge fragments in the input information and different roles in the dialogue,the pre-model is more suitable for this knowledge-fused dialogue generation task.In addition,non-likelihood loss is used as an additional auxiliary task to further improve the diversity of reply generation.And through the joint training method of curriculum learning,the error accumulation phenomenon in the training stage is alleviated,and the external knowledge consistency and model robustness of generated replies are improved.Experiments were conducted on the Chinese multi-domain dialogue dataset KdConV,and the results showed that compared with the baseline model,the proposed model in this paper increased the F1 score by an average of 6%in the movie,music,and travel domains,BLEU-2 score by 8%,and Distinct-2 score by 5%.3.A multi-turn dialogue system has been designed and implemented based on the proposed knowledge-driven multi-turn dialogue generation model proposed.Tourism domain knowledge is introduced to provide realtime and accurate travel information through multi-turn interactions,meeting user needs. |
