Code reproduction of EMNLP-2019 Paper: Enhancing Local Feature Extraction with Global Representation for Neural Text Classification

1. Data download:

   1. download data from here: https://drive.google.com/drive/u/1/folders/0Bz8a_Dbh9Qhbfll6bVpmNUtUcFdjYmF2SEpmZUZUcVNiMUw1TWN6RDV3a0JHT3kxLVhVR2M and put to data/raw_data
   2. download glove.840B.300d.txt and put to data/raw_data

2. requirements

    nltk
    tensorflow 1.4 or later
    python 2.7
   

3. Data preprocess:

   in train, split the same number samples with test as the dev dataset

    python preprocess/process_public.py -p data/raw_data/ag_news_csv -n 50000
    -n: vocab
    -p data_path
   

4. Run

    python src/run_drnn.py conf/disconnected_rnn/ag_news.config -b 128
        -b 128: batch_size
        -msl 100: max_seq_len 
        -ebi 2000: eval batch interal
        -o Adadelta: optimizer (Adam/Adadelta)
        -l 1: learning rate=1
        -r 0.1: rnn dropout rate
        -fd 0.2: fc dropout
        -et cnn: global encoder type (cnn rnn attend_rnn other_transformer transformer)
        -at same_init: interaction mode (same_init, attend_init)
        -ft cnn: feature extractor type (rnn(drnn),cnn   only avaliable in drnn, default is drnn, cnn)
        
        -ld: learning_decay
        -fee: fobid eval at the end of epoch
        -l2: use l2 regularization
   

5. details command

    command-details