We analyze the neural rationale models proposed by Lei et al. (2016) and Bastings et al. (2019), as implemented in Interpretable Neural Predictions with Differentiable Binary Variables by Bastings et al. (2019). We have copied their original repository and build upon it with data perturbation analysis. Specifically, we implement a procedure to perturb sentences of the Stanford Sentiment Treebank (SST) data set and analyze the behavior of the models on the original and perturbed test sets.

You need to have Python 3.6 or higher installed. First clone this repository.

Install all required Python packages using:

pip install -r requirements.txt

And finally download the data:

cd interpretable_predictions
./download_data_sst.sh

This will download the SST data (including filtered word embeddings).

Perturbed data and the model behavior on it is saved in data/sst/data_info.pickle, results/sst/latent_30pct/data_results.pickle, and results/sst/bernoulli_sparsity01505/data_results.pickle. To perform analysis on these, skip to the Plotting and Analysis section. To reproduce these results, continue as below.

To train the latent (CR) rationale model to select 30% of text:

python -m latent_rationale.sst.train \
  --model latent --selection 0.3 --save_path results/sst/latent_30pct

To train the Bernoulli REINFORCE (PG) model with L0 penalty weight 0.01505:

python -m latent_rationale.sst.train \
  --model rl --sparsity 0.01505 --save_path results/sst/bernoulli_sparsity01505

To perform the data perturbation, run:

python -m latent_rationale.sst.perturb

This will save the data in data/sst/data_info.pickle.

To run the latent model and get the rationale selection and prediction, run:

python -m latent_rationale.sst.predict_perturbed --ckpt results/sst/latent_30pct/

For the Bernoulli model, run:

python -m latent_rationale.sst.predict_perturbed --ckpt results/sst/bernoulli_sparsity01505/

These will save the rationale and prediction information in results/sst/latent_30pct/data_results.pickle and results/sst/bernoulli_sparsity01505/data_results.pickle for the two models, respectively.

To reconstruct the plots for the CR model, run:

python -m latent_rationale.sst.plots --ckpt results/sst/latent_30pct/

To run part of speech (POS) analysis for the CR model, run

python -m latent_rationale.sst.pos_analysis --ckpt results/sst/latent_30pct/

The perturbed data is stored as a dictionary where keys are indices (ranging from 0 to 2209, as the standard SST train/validation/test split has 2210 sentences). Each value is a dictionary with an original field, containing the original SST data instance, and a perturbed field which is a list of perturbed instances where each perturbed instance is a copy of the original instance but with one token substituted with a replacement. This is all saved in data/sst/data_info.pickle.