The Research Of Entity Relationship Extraction Of Apple Diseases And Insect Pests Based On Attention Mechanism

There are many kinds of apple diseases and pests,which seriously affect apple production.Effective control measures can promote the development of apple industry.Most authoritative information of apple diseases and pests is stored in unstructured text.Due to the lack of structured management of text data,many entity relationship information cannot be obtained during the prevention and control process.In order to obtain the structured apple disease and pest relationship data,this paper uses relation extraction technology to mine entity relationship information in a large number of unstructured apple disease and pest texts,which provides a structured data foundation for the construction of apple disease and pest knowledge graph,information retrieval platform,intelligent question answering system,etc.In view of the lack of datasets in the relation extraction of apple diseases and pests,the diverse relation categories between apple disease and pest entities,the unbalanced distribution of relation categories,and the poor performance of existing relation extraction methods in the field of apple diseases and pests,this paper constructs apple disease and pest relation extraction dataset and relation extraction model.The main contents of this paper are as follows:(1)A relation extraction dataset of apple diseases and pests was constructed.In view of the lack of relation extraction datasets in the field of apple diseases and pests,this paper collected seven authoritative books under the guidance of experts as data sources,including“Control of Apple Diseases and Pests”.The 28 relation categories between apple disease and pest entities were formulated and 20060 relation instances were annotated according to the defined relation categories,which provide a dataset for the apple disease and pest relation extraction model.Compared with other relation extraction datasets in the agricultural field,the dataset constructed in this paper realizes the diversification and fine-grained categories in the field of apple diseases and pests and can guide the disease and pest control more professionally and reasonably.(2)A relation extraction model fusing dual channels and attention mechanism was proposed.The model adopts a dual-channel mechanism and each channel contains different character embedding and Bi GRU.Different channels obtain different semantic features of the same text.The multiple semantic features obtained by dual channels are further passed through the attention mechanism to improve the relation extraction capabilities of the model.Experiments show that compared with the single-channel structure,the F1-scores of the model are increased by 2.61% and 1.19% respectively,which can better mine the text semantics.Compared with baseline models such as CNN,Bi GRU and Att-BLSTM,the F1-score of this model on the apple disease and pest dataset reaches 94.02% and has increased by 0.3%~8.81%.It performs better in the case of unbalanced distribution of relation categories.(3)A relation extraction model that integrates BERT and entity representation based on attention mechanism was proposed.The model uses BERT to obtain dynamic character vector representations and entity vector representations of text.Bi GRU is used to extract the contextual semantic features of the character vector representations.The entity vector representation is introduced into the output of the attention mechanism to enhance the influence of entity features on the model and realize the classification of relation categories.Compared with the current mainstream models CNN,BLSTM,Att-BLSTM,BERT and R-BERT,this model proposed achieves the best performance on the apple disease and pest dataset with an F1-score of 97.59%,which increases by 1.77%~12.38%.It can better cope with the phenomenon of fine-grained relation classification and unbalanced distribution of relation categories,which proves the effectiveness of the model.The performance of the model on the public datasets San Wen and Fin RE is better than that of the comparison models with F1-scores of 71.29%and 50.79%,which are improved by 3.4%~11.87% and 0.78%~9.32%,and proves the model generalization.