Research On Fine-grained Fishery Standard Information Extraction Based On Rules And Deep Learning

Fisheries standardization is the main trend in the development of digital fisheries,and achieving fisheries standardization requires the support of a fisheries standard information service system,and improving the fisheries standard service system requires extracting the knowledge contained in fisheries standard texts.Information extraction technology can extract knowledge from fisheries standard texts.Currently,fisheries standard information extraction mainly focuses on extracting coarse-grained entities.However,to improve the fisheries standard service system,it is necessary to extract comprehensive fisheries standard entities,including fine-grained entities.Since fine-grained entities in fisheries standards are mainly found in fisheries standard tables and often have overlapping relationships,it is necessary to extract overlapping relationships based on table extraction to enhance fisheries standard information extraction.To address the above-mentioned issues,this study conducts research on fine-grained fisheries standard information extraction,and the specific research tasks are as follows.(1)Combining rules and deep learning for fisheries standard table extraction methods.To solve the problem of diverse table structures and unfixed table header positions in the fishery standard text,a table information extraction method that combined Rule-Based-Matching(RBM)and AbTransformer(Absolute Transformer)was proposed.The method used an improved Transformer to extract non-regular class table information,introduced row position encoding in the position encoding module,spliced it with the feature vector to obtain table row positions to address the issue of irregular rows and columns.Furthermore,AbTransformer was combined with the rule method to enhance the accuracy and generalization ability of the model.Experimental results demonstrated that the F1 value of the RBM-AbTransformer method in this study reached 95.77%,which was 5.27% higher than the F1 value of the AbTransformer model.This indicated that the proposed method in the present paper effectively improved the overall effectiveness of information extraction from fishery standard forms.(2)An ERNIE-based method for extracting standard overlapping relationships in fisheries.The ERNIE-based Text CNN-Bi LSTM-Attention model was proposed to address the phenomenon of low extraction recall due to the existence of overlapping relationships in the standard text of fisheries.To get the hidden semantic information of overlapping entities,ERNIE was introduced,while to obtain comprehensive information of input features,Text CNN and Bi LSTM were combined to extract long and short distance features,and the Attention mechanism is introduced after Bi LSTM to increase the weight of long distance features,so that the entities have different relationship representations in different features to solve the overlapping relationship problem.The experimental results show that the F1 value of the proposed method reaches 94.33%,which is 1.81% higher than that of the BERT-Bi LSTM-Att(s)model,indicating that the method in this article can effectively extract the overlapping relationships of the fishery criteria and provide a basis for the construction of the fishery criteria system.