Research On Knowledge Reasoning Algorithm Based On Logic Rules And Deep Learning

Knowledge graph is a process of structured description of information in the way of human cognition of the world,and the process of managing,storing,mining,and visualizing the complex and disorderly diversified data.It is a research hotspot in the field of artificial intelligence and is a key to intelligent services and favorable support for applications.The knowledge graph is composed of multiple triples containing entities and relationships.Traditional knowledge graph construction techniques and data mining cannot cover all knowledge well,and it is difficult to quickly filter wrong information.Knowledge reasoning is based on Existing knowledge is a key technology to complete the completion and denoising of knowledge graphs to reason about triples with missing arbitrary elements.A high-quality knowledge graph is constructed is of great significance in applications such as intelligent retrieval and decision-making assistance.This thesis takes knowledge reasoning as the research focus,conducts in-depth study and thinking on knowledge reasoning related technologies at home and abroad,and introduces the modeling process and core ideas of representative reasoning models in detail.Through the analysis of the existing reasoning method problems,the improvement ideas for the adaptability of the algorithm and the reasoning efficiency are proposed.The main research work and innovations of this thesis are as follows:(1)The entire knowledge graph often contains a large number of triples.Performing random walks and feature extraction on the global knowledge graph makes the model reasoning cycle longer,and the knowledge is not all correct which causes the risk of noise transfer easily.This thesis improves and optimizes the reasoning efficiency and accuracy rate of the model and considers the use of entity similarity information to prejudge whether the relationship is valid,so as to constrain the path walking and improve the reasoning efficiency;in addition,a hypothesis is proposed in this thesis that the relationship learning is carried out through a multi-task-oriented framework and the reasoning results of the local knowledge graph are merged to carry out collaborative reasoning.Based on the above technology,an improved model MT-CRWA is proposed to make full use of the information in the knowledge graph and further improve the inference results.(2)The knowledge graph contains a wealth of information.Merely using entity and path information is not enough to express knowledge,and it is easy to cause knowledge errors or contradictions.In addition,every time the dynamic update of the knowledge graph is performed,a node traversal needs to be performed,which causes a waste of computing resources and time.The model is improved in thesis based on the above problems and a model TM-NNR is proposed that introduces the time sensitivity information of the relationship and the storage control unit for reasoning.On the one hand,it supports the memory and reading of the triplet content,and on the other hand,it makes the reasoning process more expressive.The ability to express is more in line with the continuous iterative process of knowledge.The generalization ability of the reasoning model is an important measurement index.Monte Carlo simulation experiments are used to train the model and verify the generalization ability and reasoning ability of the improved model in actual application scenarios.