Chinese Entity Relation Extraction Based On Neural Network

Information extraction is a very important sub-field of natural language pro-cessing,which is the process of transforming unstructured text information into struc-tured data automatically.The relationship extraction in information extraction is a key sub-task,which is an important link between the past and the future.Its task is to ex-tract the semantic relationship between entities.With the deepening of the research on relationship extraction,scholars have been innovatively proposing various advanced methods to apply to relationship extraction,and the performance of extraction is also improving.According to the number of entities,it can be divided into single entity pair relationship extraction and multi entity pair relationship extraction.This thesis mainly studies the relationship extraction between single entity pair and multi entity pair.In the research of single entity pair,aiming at the single repre-sentation method of word vector and the lack of semantic relation,a neural network relation extraction model based on dependency relation and multiple word vectors is proposed.The sentences after dependency analysis are represented by Word2vec and GloVe,and then are spliced with dependency relation respectively as the input of dual channel convolution neural network Finally,the result of relation extraction is obtained.In this thesis,we do experiments on the data sets extracted by Baidu Encyclopedia and Wikipedia.The entity results show that F1 is nearly 2.1%and 0.7%higher than CNN and CR-CNN when multiple word vectors and dependencies are used,which proves its effectiveness.In the research of multi entity pairs,the sentences are constructed into a full con-nected graph according to the multi entity pairs and entity relations in sentences.The entity represents the nodes in the graph,the context of the entity pair represents the edges between nodes,and optimizes the context representation of the entity pair.The substatement extraction module is added,and then the entity graph obtained uses the path aggregation algorithm to get the relationship representation between the entity pairs,and finally it is sent to the classifier to get the relationship extraction result.In this thesis,the experiment is carried out on the data set ace2005.The experimental re-sults show that when the substatement extraction module is added,the value of F1 is increased by nearly 1.3%compared with the original model,which proves the effec-tiveness of the substatement extraction module.Figure 28 table 13 reference 85...