Research On Relation Extraction Of Enterprise Support Policy Tex

Enterprise support policies refer to a series of documents issued by the national,provincial,municipal and county governments at all levels to support the development of enterprises,such as tax relief,capital subsidies,project approval,discount interest,and title recognition,aimed at promoting the development and expansion of small and medium-sized enterprises.The implementation of relevant policies will help improve the core competitiveness of SMEs,create more employment opportunities and promote sustainable economic development.However,as unstructured data in enterprise support policies is generally difficult to be directly recognized by computers,it is necessary to convert them into structured data that can be processed by computers through certain processing methods to facilitate the establishment of policy file big data,unified management of policy files,and sharing of data.In this context,this paper uses the relationship extraction technology to process the application conditions in the enterprise support policy text.The specific work contents are as follows:(1)The text of enterprise support policy contains a large amount of unstructured data in professional fields,and the semantic relationship between entities is complex.This paper proposes a relationship extraction model based on BERT-wwm-ext baseline model,combined with bidirectional gated cyclic neural network BiGRU and multi-head attention mechanism.The process is as follows: First,the baseline model BERT-wwm-ext is used to pre-train the text of the enterprise support policy to get the initial vector representation.Then,the bidirectional gated cyclic neural network BiGRU learns the context information of the sentence and gets the overall feature vector.In the third step,multiple attention mechanism is introduced to enhance the dependency between entities and relation words in the task of relationship extraction.Finally,the relationship of enterprise support policy texts is classified through the full connection layer.To a certain extent,the adaptability of the model to the experimental data in this paper is strengthened,and the accuracy of the relational extraction task is improved.(2)Aiming at the problem of context dependence of long text entities in enterprise support policy texts,this paper integrates the method of sequence annotation based on the relationship extraction model in Chapter 4.Firstly,the input sentence is converted into a word vector,and then passed to the BiGRU layer to complete the coding and transform the relationship extraction task into an entity recognition task.Then,the conditional random field CRF predicts the entity label,Finally,the multi head attention mechanism is used to improve the semantic relationship between entities and complete relationship extraction at the full connection layer.Experiments show that the performance of the relation extraction model combined with sequence annotation is better than other model algorithms.