Chinese Relation Extraction of Apple Diseases and Pests Based on BERT and Entity Information.

Most existing methods for Chinese relation extraction suffer from fine-grained relation categories and unbalanced category distribution in the field of apple diseases and pests. To solve above problems, we construct the AppleRE dataset, which contains 28 relation categories and 20060 relation instances with the characteristic of richer categories than existing agricultural datasets. Then, we propose a relation extraction model BE-ARE introducing BERT and entity information, in which dynamic character representations and entities that reflects the unique meaning of Chinese words are utilized to enhance the data features. The performance of BE-ARE on AppleRE achieved a precision of 98.44%, a recall of 96.75% and an F1-score of 97.59%. The F1-score is increased by 1.77%-12.38%, which outperforms the comparison models. Experimental results demonstrate the competitiveness of BE-ARE when considering fine-grained classification and unbalanced category distribution. In addition, the proposed model shows the generalization on the public datasets in distinct domains.