Abstract: 【Purpose/significance】Developing domain oriented fine-grained named entity recognition is of great significance to improve
the accuracy of text mining.Taking tomato diseases and pests as an example,this paper explores the application of deep learning tech?
nology to realize domain oriented fine-grained named entity recognition.【Method/process】In this spaper,E-books,papers and web
pages are used as data sources to label six entities,including species,diseases and pests,symptoms,time,parts and control agents,the
pre-trained vectors of BERT and CBOW are input into BiLSTM-CRF model for training,after recognition,supplementary rules were ap?
plied for entity boundary control.【Result/conclusion】 Combined with pre-trained word vector by BERT and BiLSTM-CRF,the
F-score reached 81.03% after the control of supplementary rules,which is better than other models and has a better effect in entity rec? ognition in the field of tomato pests and diseases.【Innovation/limitation】The word vector pre-trained by BERT can effectively reduce the impact of segmentation errors in tomato pest domain entities.According to the characteristics of different entities,the supplementary rules can effectively control the entity boundary and improve the recognition accuracy.But the rule supplement of this paper is only in the test part,and does not join the training process,so the overall accuracy still needs to be improved.