Research On Joint Extraction Method Of Entity And Relation Based On Deep Learning

Entity and relation extraction is one of the most important research topics in the field of information extraction.Its main task consists in extracting name entities from unstructured text and having a judge of specific entity type;In addition,it also identifies if a pair of named entities are associated and determine their type of association.The construction of the high-precision model for entity and relation extraction is not only the foundation of knowledge map,smart search,automatic question-answering and other natural language processing applications,but also a contribution to vertical fields such as finance,medical treatment and e-commerce.With the rapid development of data age and the expansion with lightning speed of network information,information machine learning which is more difficult to deal with a large amount of text data,deep learning,more conducive to entity and relation extraction model construction,has become a research hotspot nowadays because of its powerful features expression and parameter learning ability.This thesis carried out research on the entity and relation joint extraction model based on deep learning.By analyzing the shortcomings of existing models,a joint extraction model with strong expression ability and high extraction accuracy was constructed.I was able to perform the following tasks:Firstly,sort out the existing entity and relation extraction model,complete the summary and the analysis of the research status of single entity,single relation and joint extraction model,and pay attention to the defects and deficiencies of existing model.Secondly,aiming at inconsistency and ambiguity of the existing evaluation indexes,improve their definitions,find out three kinds of evaluation indexes,which is easier to evaluate and measure the model performance.Thirdly,in view of the defect in the model structure,realize a basic multi-head selection model for joint extraction,which achieves dependencies between entities and relation subtasks through parameter sharing encoding layers,and improves labeling strategy of relation submodels,make model extract entities and relations at the same time,thereby reducing the effects of accumulated error among subtasks;As for the defects of the Chinese text representation and gradient return of the basic model,a multi-head selection joint extraction model based on ELMo was proposed.ELMo dynamic embedding was used to reduce the influence of polysemous words and unregistered words.Its performance was verified by experimental simulation.Finally,point at the defects of feature expression extracted from the encoding layer and the relation score layer,a multi-head self-attention joint extraction model based on improving clauses is proposed,in which the multi-head self-attention mechanism is added on encoding layer to complete the extraction of the internal structural features of the sentence and to realize the construction of different subspace information.In the hierarchy of relations,the information of two subsentences is combined separately with the entity pair information,respectively to improve the expression ability of relationship characterization between entity pairs.Its performance was also verified by experimental simulation.