Open YanLiang1102 opened 3 years ago
The recall is off from the paper result, the reason might be the paper recall consider the "negative instance" might need to take a look.
after consider O in precision and recall, the recall is quite high in a micro way. on the dev-dataset.
processed 193229 tokens with 192575 phrases; found: 192849 phrases; correct: 175886. accuracy: 91.20%; precision: 91.20%; recall: 91.33%; FB1: 91.27 achieve: precision: 0.00%; recall: 0.00%; FB1: 0.00 0 action: precision: 81.82%; recall: 29.03%; FB1: 42.86 55 adducing: precision: 0.00%; recall: 0.00%; FB1: 0.00 0 agree_or_refuse_to_act: precision: 100.00%; recall: 27.81%; FB1: 43.52 42 aiming: precision: 0.00%; recall: 0.00%; FB1: 0.00 0 arranging: precision: 100.00%; recall: 1.50%; FB1: 2.96 3 arrest: precision: 0.00%; recall: 0.00%; FB1: 0.00 0 arriving: precision: 100.00%; recall: 2.94%; FB1: 5.71 9
without using the pretrained embedding submit using bilstm-crf submit to leadboard get this result on testing:
Micro_F1: 22.920265
Micro_Precision: 93.235294
Micro_Recall: 13.066178
Macro_F1: 7.039132
Macro_Precision: 20.702837
Macro_Recall: 4.861637
The micro-recall is very low.