Research And Implementation Of Knowledge Reasoning Technology Based On Neural Tensor Network

At present,most large-scale mainstream knowledge bases such as DBpedia and NELL mostly adopt information extraction methods to automate construction.This construction method has the advantages of less manual intervention,timely update of knowledge in the knowledge base,but also brings about the problem of knowledge base completeness.Knowledge base often facing a problem that a lack of knowledge,which greatly increases the knowledge base's incompleteness.In order to improve the quality of application which is based on knowledge base,we must enhance the completeness of the knowledge base.How to complement the missing knowledge through automation has become an important issue in the construction of knowledge base.Knowledge reasoning technology is based on the existing knowledge in the knowledge base,to make up for missing knowledge,thereby enhancing the knowledge base in the knowledge base.In this paper,some problems existing in NTN,a classical model of low-dimensional entity semantic vector,are analyzed and improved,which mainly includes some work:(1): A brief overview of some important technologies in the process of constructing knowledge graph is given.The current knowledge-based knowledge inference technology is analyzed and introduced,including the advantages and disadvantages of several kinds of knowledge reasoning technologies based on logic rules,relational paths and semantic vectors.(2): A new way of initializing entity semantic vector based on entity attribute Type classification is proposed.Compared with the method of initializing entity semantic vector randomly,the average inference precision(MAP)of multiple relationships of WordNet and Freebase is 2.7 %,3.2%,respectively.(3): In view of the high complexity of the model of neuron tensor network model,an improved inference model-hybrid neural network model is proposed,which adds a standard network layer in front of the tensor network layer to reduce the input to the tensor nerve The vector dimension of the network layer reduces the complexity of the tensor network layer.Experiments show that the improved model can reduce the running time of the model effectively under the premise of not reducing the inference accuracy.(4): The particularity of the neural tensor network model is the introduction of the Tensor's structure to characterize the interaction of entity vectors.In this paper,the effect of Tensor slices on the model accuracy is explored experimentally in the neuron tensor network model.