Research On Entity Relationship Extraction Based On Deep Learning

In today's information age,most of the information people get every day that comes from the network.But the useful information for users is scattered in the web page which are mixed with noisy data.Therefore,more and more experts and scholars pay attention to how to extract useful data from massive noise data.The objective of entity relationship extraction task is to extract the relationship between the labeled entities in the text,so as to obtain the association between the knowledge.The common method of entity relationship extraction is to use a single neural network for feature extraction,but the method can only focus on the feature information of a certain aspect,and its performance is not stable.In addition,the shortest dependency path has achieved certain effects in the extraction of entity relationship,but it only focuses on the elements that play an important role for the relationship between entities in the sentence,and cannot fully obtain the contextual characteristic information of entities.Focusing on the above two problems in the study of entity relationship extraction,this thesis carried out the following two aspects of work:1.In order to solve the problem that a single neural network can only pay attention to a certain feature of a sentence and showing different results under different input conditions,an entity relation extraction model based on ensemble learning method that is proposed.Based on the idea of ensemble learning,the model integrates two weak classifiers CNN and Bi-LSTM into the strong classifier of MLP and forms a comprehensive model.This comprehensive model can not only make full use of the advantages that CNN and Bi-LSTM pay close attention to local feature information and global feature information,but also make full use of the self-learning ability and automatic weight allocation of MLP that can be used to improve the performance of relation classification.When the model is not added other additional feature information,the F1 value obtained is87.7%.2.The shortest dependency path method mainly focuses on the components which are important to the entity words in the sentence,but ignores the context characteristic information of the entity words.Therefore,this thesis proposes a method that enhances the dependency path for the extraction study of entity relations.The method mainly uses Tree-LSTM to construct the dependency subtree for the words on the shortest dependency path.In this way the subsequent model can not only pay attention to the words that have an important role on the entity words,but also provide the dependency subtree for the words on the dependency path,and fully obtain the context information of the words.Compared with other mainstream relational extraction models,the model achieves an effect of 88.79%.In order to facilitate the application of the above research content in the actual scene,this thesis designed and developed a prototype system of entity relationship extraction.