Code used for sentence classification tasks. We evaluate CNN, LSTM and SRU on 6 benchmarks. Example learning curves:

Training time (x-axis) vs valid accuracies (y-axis) on classification benchmarks

• Download the datasets from harvardnlp/sent-conv-torch/data

• Download pre-trained word embeddings such as word2vec; make it into text format

• Make sure CUDA library path and cuda_functional.py is available to python. For example,

  export LD_LIBRARY_PATH=/usr/local/cuda/lib64
  export PYTHONPATH=../../sru/

• Run train_classifier.py and get the results.

  python train_classifier.py --help           # see all running options

  python train_classifier.py --dataset mr     # which dataset (mr, subj, cr, sst, trec, mpqa) 
        --path data_directory                 # path to the data directory
        --embedding google_word2vec.txt       # path to pre-trained embeddings
        --cv 0                                # 10-fold cross-validation, use split 0 as the test set

Part of the code (such as text preprocessing) is taken from https://github.com/harvardnlp/sent-conv-torch

CNN model is the implementation of (Kim, 2014), following

• torch / lua version: https://github.com/yoonkim/CNN_sentence
• pytorch version: https://github.com/Shawn1993/cnn-text-classification-pytorch