Key Technology Study And Model Realization Of Knowledge Grounded Dialog System

In recent years,deep learning has achieved great success in natural language processing,and dialog systems attract much attention as a significant application in natural language processing.For a long time,researchers have been committed to building intelligent robots that can naturally talk to humans,and how to endow robots with the ability to recognize and apply knowledge information is one of the key challenges in building intelligent robots.Recently,Knowledge grounded dialog systems usually use the external knowledge base,knowledge graph,or additional textual information as knowledge input for dialog systems.However,the scale of external knowledge is large in real application scenarios,and it is difficult to select appropriate knowledge as the input of dialog effectively.In addition,knowledge grounded dialog datasets have relatively small sizes.The system is easy to overfit to the given knowledge information in these datasets during the training phase,resulting in poor performance in scenarios with new knowledge.To address the above problems,we carry out the following research work in this paper:(1)Multi-subgoal Driven Recommender System Based on Knowledge-Enhanced Framework: We first take recommendation dialog with multiple subgoals task an example and propose a Knowledge-enhanced Multi-subgoal Driven Recommender System(KERS)to efficiently select fine-grained knowledge from the knowledge graph and integrate it into the recommendation dialog system.KERS uses a dialog guidance module to plan a series of dialog subgoals and select fine-grained knowledge for the current subgoal.Moreover,we propose a sequential attention mechanism,a noise filter,and a knowledge enhancement module to incorporate the input knowledge in responses.The experimental results on the Du Rec Dial dataset demonstrate that KERS has better knowledge utilization ability than the previous models.(2)Robust knowledge grounded dialog system: In order to further improve the generalization ability of knowledge grounded dialog systems,we propose a general Robust Knowledge Grounded Dialog System(Ro KGDS).Ro KGDS uses a bucket encoder to uniformly and efficiently encode structured and unstructured knowledge and uses a hybrid attention mechanism to effectively combine pre-trained language models to improve model robustness.Additionally,to effectively use the input knowledge,we use a copy mechanism to directly copy the information from the input knowledge to the responses.We conduct extensive experiments on the Du Conv,Du Rec Dial,and CPC datasets,and the results show that Ro KGDS is more robust than other models.(3)Knowledge Grounded Dialog System Based on Entity Perturbation: An important reason for the system's poor performance in scenarios with new knowledge is that the system generates factually wrong responses – hallucination problem.We visualize the model internal attention matrix for the hallucination and propose a Learning Hallucination problem.Learning hallucination is the problem that the model learns incorrect knowledge mapping during the training process,which is one of the important reasons for producing hallucination.In this paper,we propose an Entity Perturbation(En PT)method to adjust the input knowledge to effectively alleviate the problem of learning hallucinations.Experiments on DuConv and Du Recdial demonstrate the effectiveness of EnPT.