Knowledge Fusion Via Tensor Decomposition And Its Application In Dialogue System

The rapid development of the Internet has brought the explosive growth of information,resulting in its chaotic structure.For this reason,the knowledge graph is introduced and has achieved great development.Nowadays,knowledge graphs are widely used in search engines,recommendation systems,intelligent question answering,dialogue systems,etc.This paper mainly focuses on the applications of knowledge graph in task-oriented dialogue system.In practice,a dialogue system often covers knowledge in different sources correspond-ing to different fields separately.These knowledge bases from different sources are often maintained by skilled personnel and have heterogeneous distributions and attributes.Such a situation makes it difficult or even impossible for the knowledge base to be directly applied to task-oriented dialogue systems.Thus,we exploit the knowledge fusion model to fuse these heterogeneous knowledge bases into one knowledge base,and then apply the fused knowl-edge graph to the task-oriented dialogue system in which user's intent recognition plays a very important role.Understanding the user's intent accurately can speed up the user's prob-lem resolution and improve the user's conversation experience.However,recent studies utilize deep neural networks as intent recognition classifiers.Deep neural network is a black box and lacks interpretability.The knowledge graph is rich in structured knowledge,which offers possibilities for the interpretability of the intention recognition model.As a result,knowledge fusion and graph-based intent recognition have become two crucial technologies in task-oriented dialogue system based on knowledge graph.In detail,knowledge fusion is the theoretical basis for constructing a task-oriented dialogue system based on knowledge graph.User intent recognition is an integral part of task-oriented conversations.Therefore,this paper mainly researches on knowledge fusion and user's intent recognition based on knowledge graph in the task-oriented dialogue system.First,this paper takes the fusion of knowledge graphs pertaining to the part-of relation as an example,and studies the fusion of knowledge graphs with large amounts of completely separate connected components and no overlap between the training and test set entities.In order to solve this problem,we introduce similarity matrix as auxiliary information and propose a model that jointly factorizes triples tensor and similarity matrix.Due to the inde-pendence between relations in the knowledge graph,the model uses the RESCAL method to decompose the triples tensor.After that,we use the Alternating Direction Method of Multi-pliers(ADMM)to optimize the model.In the experiments,by comparing with the RESCAL model and the TransE-style models,the model proposed in this paper shows the best results in accuracy,and successfully solves the fusion of knowledge graphs with large amounts of completely separate connected components and no overlap between the training and test set entities.With the help of knowledge fusion technology,we can fuse knowledge bases from mul-tiple fields into one knowledge base and apply it to a task-oriented dialogue system.In this system,knowledge graphs can help the system to quickly and accurately identify the user's intentions,and make the system highly interpretable.Therefore,this paper proposes a user intent recognition model based on intent knowledge graph.This model converts the fused multi-domain knowledge base into a intent knowledge graph representing user's intents,and relies on the reinforcement learning method to reason in the knowledge graph to obtain a suitable path.The last node in the path is the user's intent.In the experiment,compared with the model of supervised learning,the model shows a high accuracy of intent recognition and even has a high interpretability,which helps us to quickly understand the error examples generated by the model and put forward methods to improve the performance in time.In summary,this paper focuses on the two key technologies of knowledge graph in task-oriented dialogue system.These two technologies are knowledge fusion and user's intention recognition based on knowledge graph.Among them,knowledge fusion is the theoretical basis for constructing a dialogue system based on knowledge graph.In order to use multi-domain knowledge bases in task-oriented dialogues,we use knowledge fusion technology to fuse multi-domain knowledge bases into one knowledge base.User's intention recognition is an integral part of task-based dialogue.The application of the knowledge graph makes the user's intention recognition model highly interpretable.Firstly,we propose a model that jointly decomposes tensor and matrix to fuse a knowledge graph with large amounts of completely separate connected components and no overlap between the training and test set entities.After that,we propose a user's intent recognition model based on the fused knowledge graph.The model utilizes reinforcement learning to generate reasonable paths that correspond to the user's intents.Both two models have good experimental performances However,they still have a lot of room for improvement.For example,some error examples generated by the knowledge fusion model are against to common sense;the intent recognition model is still not as accurate as the supervised learning model.These problems need our further research.