conda create -n contrastive python=3.8
conda activate contrastive
pip install allennlp==1.2.0rc1
pip install allennlp-models==1.2.0rc1.dev20201014
pip install jupyterlab
pip install pandas
bash scripts/download_data.sh

If you're here just to know how we implemented contrastive projection, here it is:

u = classifier_w[fact_idx] - classifier_w[foil_idx]
contrastive_projection = np.outer(u, u) / np.dot(u, u)

Very simple :)

contrastive_projection is a projection matrix that projects the model's latent representation of some example h into the direction of h that separates the logits of the fact and foil.

bash scripts/train_sequence_classification.sh 

Run the notebooks/mnli-highlight-featurerank.ipynb or notebooks/bios-highlight-featurerank.ipynb jupyter notebooks.

These notebooks load the respective models, and then run the highlight ranking procedure.

First, cache the model's encodings of the dev set examples:

bash scripts/cache_encodings_bios.sh

Then run the notebooks/bios-highlight-foilrank.ipynb notebook.

First, cache the model's encodings of the dev set examples (skip if already executed):

bash scripts/cache_encodings_bios.sh

Then run the notebooks/bios-foilpower.ipynb notebook.

First, generate concept labels as a numpy matrix from the BIOS dataset:

python scripts/bios_concepts.py --data-path data/bios/train.jsonl --concept-path experiments/models/bios/roberta-large/concepts/gender-male/train
python scripts/bios_concepts.py --data-path data/bios/dev.jsonl --concept-path experiments/models/bios/roberta-large/concepts/gender-male/dev
python scripts/bios_concepts.py --data-path data/bios/test.jsonl --concept-path experiments/models/bios/roberta-large/concepts/gender-male/test

Then, run Amnesic Probing:

• WIP - to be added soon. Alternatively, refer to the original amnesic probing repository which has the necessary code.

First, generate concept labels as a numpy matrix from the BIOS dataset:

python scripts/mnli_concepts.py --data-path data/mnli/train.jsonl --concept-path experiments/models/mnli/roberta-large/concepts/overlap/train
python scripts/mnli_concepts.py --data-path data/mnli/dev.jsonl --concept-path experiments/models/mnli/roberta-large/concepts/overlap/dev
python scripts/mnli_concepts.py --data-path data/mnli/test.jsonl --concept-path experiments/models/mnli/roberta-large/concepts/overlap/test

Then, run Amnesic Probing:

• WIP - to be added soon. Alternatively, refer to the original amnesic probing repository which has the necessary code.

The examples we used for the negation concept analysis are:

data/nli_negation_concept/entailment.jsonl  # entailment instances
data/nli_negation_concept/entailment_with_negation.jsonl  # the above entailment instances, paraphrased with negation words
data/nli_negation_concept/neutral.jsonl  # neutral instances
data/nli_negation_concept/neutral_with_negation.jsonl  # the above neutral instances, paraphrased with negation words

To analyze them with respect to the trained MultiNLI model, run the notebook notebooks/mnli-negation-foilrank.ipynb.