1024er / MLM_transfer

Implemetation of MLM_transfer

MLM_transfer

Implemetation of MLM_transfer

Environment：

• python==3.6
• pytorch==0.4.1
• theano==1.0.4
• nltk==3.0.0b2 (included)

Procedures：

Mask stage

• Three methods：
  • attention_based method
  • frequency_ratio method
  • fusion_method method
• Command（Results already included, no need to run it again） bash run_preprocess.sh

Fill stage

• Two steps：

  • MLM -> fine_tune_cbert.py
  • MLM-SS -> fine_tune_cbert_w_cls.py
  • MLM-PG -> fine_tune_cbert_w_cls_pg.py

• Commands

  • Corresponds to attention_based mask method

    • bash scripts/attention_based/fine_tune_yelp_attention_based.sh
    • bash scripts/attention_based/fine_tune_amazon_attention_based.sh

  • Corresponds to frequency_ratio mask method

    • bash scripts/frequency_ratio/fine_tune_yelp_frequency_ratio.sh
    • bash scripts/frequency_ratio/fine_tune_amazon_frequency_ratio.sh

  • Corresponds to fusion_method mask method

    • bash scripts/fusion_method/fine_tune_yelp_fusion_method.sh
    • bash scripts/fusion_method/fine_tune_amazon_fusion_method.sh

Note: The accuracy results produced here are lower than original paper, but the BLEU scores are higher. It is a trade-off between accuracy and BLEU. To achieve the same results from paper, you just need to modify fine_tune_cbert_w_cls.py: if lm_loss.item() > 1.5: => if lm_loss.item() > 1.7 or higher # line 153

We also tried to use policy_gradient instead of soft-sampling to back-propagate gradient, and we encourage you to implement it yourself.