Research And Application Of Relation Classification Technology Based On Deep Learning

In recent years,many large-scale knowledge bases have been widely used in natural language processing,web search and automatic question and answer and other aspects.The core of these knowledge bases is the factual tuple,like "(Mark Zuckerberg,the founder of Facebook)".However,these knowledge bases are far from being enough to describe the knowledge in the real world.To further enrich the existing knowledge bases,recent researchers hope to improve the volume of the knowledge base through the automated approach,especially to increase the number of factual relations.This process is called relation extraction,that is,identify and generate the semantic relationship between entities from the unformatted text.Relation classification,classifying the relation of entity pairs to obtain semantic relationship,is an implementation method of relational extraction.This paper first introduces the research about relation extraction and neural network based relation classification at home and abroad finds that the Convolutional Neural Network(CNN)based methods,using one standard convolution layer,pool layer,and softmax layer,achieve competitive performance with other complex-structured networks.However,the traditional CNN based method ignores the fact that the words between entities are highly correlated with their semantic relations,and these methods can't obtain high-level features with only one convolution layer.Aimed at these problems,this paper proposes a simple model: the sentence is divided into three segments according to the entities,and the hierarchical Convolutional Neural Network is used to extract the high-level features of the sentence.Experiments show that the method has improved the traditional method and find two problems: when handling samples with long distance between entities,however,CNN fails to extract effective features and even extracts the wrong ones from clauses,which results in downgrades of accuracy.Besides,it is observed that existing methods produce inconsistent results when fed with forward and backward instances of the same sample.Aimed at these problems,this paper proposes selective attention based sentence encoder and bidirectional instances relation classification framework(SA-CNN + BDI).The implicit features of the keywords,words in the shortest dependent path between the two entities,are enhanced by the selective attention layer.Besides,the decision fusion strategies are used to combine the relations of forward and reverse instance to avoid the conflict.The experiment proves the validity of the two parts of the model,and the model provides state-of-the-art performance.Finally,in the background of human intelligence analysis,this paper builds a prototype system with the real-world data set to show the application in the relation extraction and knowledge graph construction.The prototype system realized in this paper can automatically extract the human relation in the Chinese text and display it through the visual interface.