GCN Relation Extraction Method Integrating Degree Information, Semantic Position Attention And Dual Type Embeddin

Entity relation extraction is one of the core technologies in natural language processing.It can help us analyze and extract the entities in sentences better,and can help us infer the semantic relations between them more accurately.In recent years,many scholars have begun to explore the joint extraction method to carry out the research of relation extraction.In the current research,joint extraction has avoided the problems of error propagation and interaction loss in the pipeline relation extraction method to some extent,but it still can’t solve the problems of dealing with overlapping relationships,insufficient semantic feature representation and insufficient model dependence.To solve the above problems,this paper proposes a Graph Convolutional Networks(GCN)relation extraction method based on double-type embedding and semantic position attention of fusion degree information,which improves the vector representation of sentences by using the position information and auxiliary information contained in sentences,and then improves the learning ability of the model through GCN to obtain more advanced relation extraction performance.This paper’s primary focus is as follows:Construct the semantic position attention mechanism of fusion degree information.Syntactic dependency tree can clearly express the logical relation between words in a sentence.therefore,the relative position distance of each word in a sentence to a given entity word on the syntactic dependency tree is extracted to construct a position vector,and then the influence weight of words is calculated to judge the importance of words.on this basis,the implicit degree information in the syntax dependency tree is combined to take into account the influence of different nodes on connecting edges.then the influence of the improved dependency tree and semantic position is considered together to further achieve the purpose of noise reduction.Facing the problem of insufficient semantic features caused by long-tail,which is mostly ignored by existing relationship extraction methods,this paper proposes to aggregate the implicit high-order feature information of similar sentences by using the aggregation of GCN and the similarity between sentences,and boost the performance of the data-poor classes at the tail through the knowledge from data-rich classes at the head of the distribution,so as to boost the model’s cognitive ability and further enhance the efficiency of relation extraction.The input vector is optimized by using the entity type information and relation type information contained in the sentence.In this paper,the entity type information and relation tag information are merged into the sentence embedding.An entity may have multiple types of tags in a sentence,and not all relational tags are equally important to the words in the sentence.therefore,it is necessary to select the most likely entity type tags and distinguish the importance of different relational tags to each word in the sentence through the type-aware attention mechanism.It can not only improve the accuracy of relation extraction,but also capture the interaction between entities and relations.The model proposed in this paper is verified on the datasets New York Times(NYT)and Web NLG,and the F1 values on NYT and Web NLG are 92.4% and91.1% respectively,which is the best among all the comparison models.The results show that the proposed model can effectively solve the problems in relation extraction and improve the correctness of relation extraction.