Parse Chinese company name to different parts correctly. For example, the name 大连泰和道农业发展有限公司 will be parsed to 大连 / 泰和道 / 农业发展 / 有限公司 with label LOC / NAME / INDU / ORG.
The predict result is as follow:
[武汉]【翔海】{水处理}<有限公司>
[漯河市]【天利】{白瓜籽加工}<有限责任公司>
[大连]【泰和道】{农业发展}<有限公司>
[天津]【大致】【和堂】{医药}<有限公司>【延年】{中药}<店>
With bert-base-chinese, the result is as follows on test set:
precision recall f1-score support
INDU 0.98 0.98 0.98 200
LOC 1.00 0.99 1.00 199
NAME 1.00 1.00 1.00 204
ORG 0.98 0.99 0.98 201
micro avg 0.99 0.99 0.99 804
macro avg 0.99 0.99 0.99 804
weighted avg 0.99 0.99 0.99 804
To reproduce:
sequence_labeling_company_train.ipynb
To predict:
sequence_labeling_company_predict.ipynb
Key ideas to get this working are due to this github.