a b/bilstm_crf_ner/README.md 

  1
  
# BiLSTM + CRF for NER





  
  

  2
  
Before running anything, make sure you have generated the pre-processed dataset using the generate_data.py file with the command mentioned in the parent directory. 





  
  

  3
  





  
  

  4
  
## Additional Requirements





  
  

  5
  
1.  Packages: Anaconda, Pytorch, AllenNLP  





  
  

  6
  
2.  Glove 300B embeddings





  
  

  7
  





  
  

  8
  
## Training





  
  

  9
  
1. Change settings in model/config.py  





  
  

  10
  
2. Main settings to change: File directories, model hyperparameters etc.  





  
  

  11
  
3. Run build_data.py  





  
  

  12
  
    - Builds embedding dictionary, text file of words, chars tags, as well as idx to word and idx to char mapping for the model to read  





  
  

  13
  
        





  
  

  14
  
4. Run train.py  





  
  

  15
  





  
  

  16
  
## Results





  
  

  17
  
|              |precision|    recall|  f1-score|





  
  

  18
  
|:---:|:---:|:---:|:---:|





  
  

  19
  
|         ADE |      0.1807 |     0.7168 |     0.2887  |





  
  

  20
  
|         Dosage |      0.9272 |     0.9123 |     0.9197  |





  
  

  21
  
|        Drug |      0.8898 |     0.9287|     0.9088  |





  
  

  22
  
|         Duration |      0.8882 |     0.7778 |     0.8293  |





  
  

  23
  
|         Form |      0.9840 |     0.9172 |     0.9494  |





  
  

  24
  
|         Frequency |      0.9412 |     0.9494 |     0.9453  |





  
  

  25
  
|         Reason |      0.7883 |     0.5238 |     0.6294  |





  
  

  26
  
|         Route |      0.9583 |     0.9226 |     0.9401  |





  
  

  27
  
|         Strength |      0.9769 |     0.9683 |     0.9726  |





  
  

  28
  
|   micro avg |      0.8708 |     0.8957 |     0.8831  |





  
  

  29
  
|   macro avg |      0.8624 |     0.8792 |     0.8684  |