CS224u Natural Language Understanding

CS224U Final Project (Spring 2019)

Setup the Environment

1. Follow instructions for setting up the CS224U environment
2. Run conda install allennlp -c pytorch -c allennlp -c conda-forge
3. Run pip install allennlp

Setup the Adversarial Datasets

1. Download the ADDAMOD and SUBOBJSWAP data sets from: https://www.dropbox.com/s/lev4327zn833dle/SharingDataResource.zip?dl=0
2. Download the "Breaking SNLI" word-replacement data set from: https://github.com/BIU-NLP/Breaking_NLI
3. Setup this structure in your project directory:

/path/to/project_home/data/

• snli_1.0
  • snli_1.0_dev.jsonl
  • etc
  • etc
• snli_adversaries
  • ADDAMOD <-- from Step 1 above
    • add_amod(dev).tsv
    • add_amod(train).tsv
  • SUBOBJSWAP <-- from Step 1 above
    • sub_obj_swap(dev).tsv
    • sub_obj_swap(train).tsv
  • breaking_README.txt
  • SUB_ADD_README.txt
  • dataset.jsonl <-- from Step 2 above

Pre-Processing

1. For faster training and testing, it is advised that you run the preprocessing.py script on the data. This parses the sentences for their SRL tags in advance.

Setup the LSTM GloVe + SRL Model

1. Activate the cs224u standard virtual environment
2. Run jupyter notebook adversarial_snli.ipynb
3. Modify the filepaths as needed, in particular the GLOVE_HOME, and near the bottom of the notebook, the various training and evaluation runs which point to specific data files.
4. Run all cells.

Baseline Model

1. Modify the code such that glove.5B.100d.txt is used for the glove_lookup
2. Alter the glove_srl_phi function to return GloVe lookup tensors only, not the concatenated GloVe + SRL tensors.