Skip to content

Latest commit

 

History

History
13 lines (12 loc) · 731 Bytes

trf-ptb.md

File metadata and controls

13 lines (12 loc) · 731 Bytes

LMs trained on PTB and evaulated by rescoring the 1000-best of WSJ'92 test set

  • run_baseline_ngram.py: train ngram LMs and used to rescore n-best list. This need install SRILM toolkit first by
cd tools
./install_srilm.sh
  • run_lstm.py: train LSTM LM with standard softmax. The source code of LSTM is in tfcode/lm/lstmlm.py.
  • run_lstmlm_nce.py: train LSTM LM using BNCE.
  • run_trf_<feat_type>_<train_method>.py': train TRF LMs with <feat_type>using<train_method>`
    • <feat_type> = discrete (discrete features) or neural (neural network features)
    • <train_method> = sa (AugSA + JSA training) or nce (NCE training)
    • for DNCE training, see egs/ptb_fake_nbest/run_trf_neural_nce.py