Research On Knowledge Extraction Method For Chinese Knowledge Graph Construction

With the development of the Internet,network data has begun to show a spurt of growth,and it is difficult for traditional management methods to effectively process this massive amount of information.As a result,the research on the construction of the knowledge graph has gradually developed.By constructing the knowledge graph,the information can be effectively managed,organized,and used.Knowledge extraction is the core part of the knowledge graph construction process,and the effect of knowledge extraction directly affects the quality of the knowledge graph.However,the current research on knowledge extraction still has problems such as low entity extraction quality and poor relation extraction effect.Therefore,this thesis focuses on the two techniques of entity extraction and relation extraction in knowledge extraction.1.In the process of extracting Chinese entities based on deep neural networks,traditional methods have problems such as character vectors cannot represent polysemous characters,character-based models cannot use vocabulary information,and most methods do not adequately use sentence internal features.In response to the above problems,this thesis proposes a Chinese entity extraction method based on the BERT-FLAT-Attention-CRF model.Firstly,use the BERT pre-training language model to dynamically generate the vector sequence that can represent polysemous characters;secondly,use the Flat-lattice model to fully fuse the vocabulary information in the vector sequence;next,use the attention mechanism to further extract the internal features of the sequence;finally,output sequence annotation results through CRF model.The method proposed in this thesis is tested on three data sets of MSRA,Weibo and Resume,and the F values are 96.15%,71.80% and 96.38% respectively,which is better than other comparison methods.2.Chinese relation extraction methods based on deep neural networks have greatly improved the performance of relation extraction compared with traditional methods,but these methods generally have the problems that the semantic vector encoded by the model cannot represent the polysemy of the Chinese sentence,the structure of the model is complex,and the internal information of the sentence cannot be fully mined.In response to the above problems,this thesis proposes a Chinese relation extraction model based on the BERT-Bi GRU-Attention.Firstly,the input sentence is encoded by the BERT pre-training language model to generate the semantic vector sequence that can represent Chinese polysemous characters;secondly,the semantic vector sequence is input into the Bi GRU model which has the relatively simple structure for feature extraction to generate the feature vector sequence;next,the attention mechanism is used to further mine the internal information of the feature vector sequence;finally,the relation classification results are output through the fully connected layer and Softmax.The experimental results show that the method can improve the effect of Chinese relation extraction while simplifying the complexity.