Note: This project is mostly based on https://github.com/yuhaozhang/sentence-convnet

• Python 2.7
• Tensorflow (tested with version 0.10.0rc0)
• Numpy
• Scipy

To visualize the results (visualize.ipynb)

• Pandas
• Matplotlib
• Scikit-learn

• We used Penn Discourse Treebank ver. 2.0.
  Assume that CoNLL 2016 data is stored in json format under data/conll dir.

  cnn-al-tf
  ├── ...
  ├── word2vec
  └── data
      └── conll
          ├── pdtb-dev.json
          ├── pdtb-dev-parses.json
          ├── pdtb-train.json
          └── pdtb-train-parses.json
  

• word2vec directory is empty. Please download the Google News pretrained vector data from this Google Drive link, and unzip it to the directory. It will be a .bin file.

python ./util.py

It creates vocab.txt, *.ids and emb.npy files.

• Hierarchical multi-label classification with negative sampling (HML+NS):

  python ./train.py --sent_len=163 --vocab_size=34368 --num_classes=21 \
  --hierarchical=True --negative=True --use_pretrain=True

• Hierarchical multi-label classification on split contexts with negative sampling (HML+NS+Split):

  python ./train_split.py --sent_len=100 --vocab_size=34368 --num_classes=21 \
  --hierarchical=True --negative=True --use_pretrain=True

• Active learning with max-entropy strategy:

  python ./train_active.py --sent_len=163 --vocab_size=34368 --num_classes=21 \
  --hierarchical=True --negative=False --use_pretrain=True --strategy=max_entropy \
  --num_epochs=1 --batch_size=100 --pool_size=1000 --log_step=1 --summary_step=10

Caution: A wrong value for input-data-dependent options (sent_len, vocab_size and num_classes) may cause an error. If you want to train the model on another dataset, please check these values.

• Display F1 and AUC score (overall performance)

  python ./eval.py --checkpoint_dir=./train/1473898241

• Display classification report (class-wise performance)

  python ./predict.py --checkpoint_dir=./train/1473898241

Replace --checkpoint_dir with the output from the training.

tensorboard --logdir=./train/1473898241

• http://github.com/yuhaozhang/sentence-convnet
• http://github.com/dennybritz/cnn-text-classification-tf
• http://www.wildml.com/2015/12/implementing-a-cnn-for-text-classification-in-tensorflow/
• http://tkengo.github.io/blog/2016/03/14/text-classification-by-cnn/
• Mihaylov and Frank. Discourse Relation Sense Classification Using Cross-argument Semantic Similarity Based on Word Embeddings ACL 2016
• Nguyen and Grishman. Relation Extraction: Perspective from Convolutional Neural Networks NAACL 2015
• Zhang and Wallace. Active Discriminative Word Embedding Learning arXiv:1606.04212