Skip to content

Latest commit

 

History

History
37 lines (30 loc) · 941 Bytes

README.md

File metadata and controls

37 lines (30 loc) · 941 Bytes

bert_Bilstm_Crf

This model based on google's Bert and hanxiao's bert-as-service this just a try to use bert in a word-embedding mode.

This model still has some problems need to solve. such as:

  1. the bert-service can catch the speed of model training.
  2. should i use trainable=True for each embeddings in each sentence
  3. The result on dev data was not very perfect!
  4. only trained 9177 batch. if more maybe better!
  5. others

Run: First:start bert service

bert-serving-start -max_seq_len 250 -pooling_strategy NONE -pooling_layer -4 -3 -2 -1 -model_dir /home/kyzhou/bert/uncased_L-12_H-768_A-12 -num_wor 4

Second:

python train.py

Third:

python eval.py

Result: loss:

on dev set: