This repo is still under construction. More information (including the paper) will be released by 1st April, 2019.

Paper link: Pending

1. Abstract
2. Code
3. Acknowledgement
4. Citation

cd src_reject
sh run.sh

In order to run the code, please check the following issues.

• Package dependencies:
  • Python 3.5
  • TensorFlow 1.11.0
  • TensorLayer 1.11.0
  • Numpy 1.14.5
  • Pandas 0.21.0
  • NLTK 3.2.5
• Download original dataset
  • GloVe.6B.200d
  • ConceptNet v5.6.0
  • DBpedia ontology dataset
  • 20 Newsgroups original 19997 docs
• Check config.py and update the locations of data files accordingly. The config.py also defines the locations of intermediate files.
• The intermediate files already provided in this repo
  • classLabelsDBpedia.csv: A summary of classes in DBpedia and linked nodes in ConceptNet.
  • classLabels20news.csv: A summary of classes in 20news and linked nodes in ConceptNet.
  • Random selection of seen/unseen classes in DBpedia with unseen rate 0.25 and 0.5.
  • Random selection of seen/unseen classes in 20news with unseen rate 0.25 and 0.5.
  • Note: seen/unseen classes were randomly selected for 10 times. You may randomly generate another 10 groups of seen/unseen classes.
• The intermediate files need to be manually generated
  • Appropriate preprocessing is recommended. For example, the vocabulary is limited by 20K most frequent words and all numbers are excluded.
  • Run combine_zhang15_dbpedia_train_test() in playground.py:
    • The generated full.csv is used to create vocabulary for DBpedia later.
  • Run doing_sth_on_20_news() in playground.py:
    • This function automatically collects 20news data and randomly splits the data into training set train.csv (70%) and testing set test.csv (30%).
    • Besides, full.csv is also generated and is used to create vocabulary for 20news later.
    • Note that the variable home_dir in this function should be the location of the home directory of uncompressed 20news data, which includes a collection of folders named by class labels.
• Other intermediate files should be generated automatically when they are needed.

Pending

An example:

python3 train_seen.py \
        --data dbpedia \
        --unseen 0.5 \
        --model vw \
        --sepoch 1 \
        --train 1

The arguments of the command represent

• data: Dataset, either dbpedia or 20news.
• unseen: Rate of unseen classes, either 0.25 or 0.5.
• model: The model to be trained. This argument can only be
  • vw: the inputs are embedding of words (from text)
• sepoch: Repeat training of each epoch for several times. The ratio of positive/negative samples and learning rate will keep consistent in one epoch no matter how many times the epoch is repeated.
• train: In Phase 1, this argument does not affect the program. The program will run training and testing together.
• rgidx: Optional, Random group starting index: e.g. if 5, the training will start from the 5th random group, by default 1. This argument is used when the program is accidentally interrupted.
• gpu: Optional, GPU occupation percentage, by default 1.0, which means full occupation of available GPUs.
• baseepoch: Optional, you may want to specify which epoch to test.

An example:

python3 train_unseen.py \
        --data 20news \
        --unseen 0.5 \
        --model vwvcvkg \
        --ns 2 --ni 2 --sepoch 10 \
        --rgidx 1 --train 1

The arguments of the command represent

• data: Dataset, either dbpedia or 20news.
• unseen: Rate of unseen classes, either 0.25 or 0.5.
• model: The model to be trained. This argument can be (correspond with Table 6 in the paper)
  • kgonly: the inputs are the relationship vectors which are extracted from knowledge graph (KG).
  • vcvkg: the inputs contain the embedding of class labels and the relationship vectors.
  • vwvkg: the inputs contain the embedding of words (from text) and the relationship vectors.
  • vwvc: the inputs contain the embedding of words and class labels.
  • vwvcvkg: all three kinds of inputs mentioned above.
• train: 1 for training, 0 for testing.
• sepoch: Repeat training of each epoch for several times. The ratio of positive/negative samples and learning rate will keep consistent in one epoch no matter how many times the epoch is repeated.
• ns: Optional, Integer, the ratio of positive and negative samples, the higher the more negative samples, by default 2.
• ni: Optional, Integer, the speed of increasing negative samples during training per epoch, by default 2.
• rgidx: Optional, Random group starting index: e.g. if 5, the training will start from the 5th random group, by default 1. This argument is used when the program is accidentally interrupted.
• gpu: Optional, GPU occupation percentage, by default 1.0, which means full occupation of available GPUs.
• baseepoch: Optional, you may want to specify which epoch to test.