This repo contains code for the paper Improving Interpretability via Explicit Word Interaction Graph Layer published in AAAI 2023.
https://arxiv.org/abs/2302.02016
Recent NLP literature has seen growing interest in improving model interpretability. Along this direction, we propose a trainable neural network layer that learns a global interaction graph between words and then selects more informative words using the learned word interactions. Our layer, we call WIGRAPH, can plug into any neural network-based NLP text classifiers right after its word embedding layer. Adding the WIGRAPH layer substantially improves NLP models' interpretability and enhances models' prediction performance at the same time.
@article{sekhon2023improving, title={Improving Interpretability via Explicit Word Interaction Graph Layer}, author={Sekhon, Arshdeep and Chen, Hanjie and Shrivastava, Aman and Wang, Zhe and Ji, Yangfeng and Qi, Yanjun}, journal={arXiv preprint arXiv:2302.02016}, year={2023} }
- Pytorch==1.5.1
- Transformers==3.3.0
- datasets=1.1.2
The data folder contains the dataset that are used for WIGRAPH.
The datasets used can be downloaded from : datasets link. To run the code, save these in data folder.
The metadata folder contains precomputed counts of words and interactions and are used to extract a subset of words and interactions used to learn the model. These can be downloaded from metadata link(precomputed counts). After downloading, save these in metadata.
The models used here as classifiers can be downloaded from : finetuned models. After downloading, save these in finedtuned_models/
The wigraph layer that can be plugged in front of any NLP classifier can be found in imask.py.
To train a model using WIGRAPH:
main.py --per_gpu_train_batch_size 32 --per_gpu_eval_batch_size 64 --task-name sst2 --learning_rate 1e-05 --factor 1000.0 --mask-hidden-dim 32 --backbone distilbert --save-dir results/distilbert/ --init-mode static --beta-i 0.0 --beta-g 1.0 --beta-s 1.0 --imask-dropout 0.3 --project WIGRAPH --non-linearity gelu --num_train_epochs 10 --seed 42 --onlyA --max_sent_len 56 --anneal
The relevant hyperparameters are
- task-name : [sst2, sst1, trec, subj, imdb, agnews]
- backbone : [distilbert, roberta, bert]
- beta-i : 0.0 for WIGRAPH-A, 1.0 for WIGRAPH-A-R
- beta-s : sparsity weight
- onlyA : does not train R, enable for WIGRAPH-A
- max_sent_len : [sst2(56), sst1(56), trec(32), subj(128), imdb(512), agnews(128)]
- We provide interpretation using SHAP as well as LIME. A WIGRAPH model can be interpreted using the following:
interpret.py --per_gpu_train_batch_size 32 --per_gpu_eval_batch_size 64 --task-name sst2 --learning_rate 1e-05 --factor 1000.0 --mask-hidden-dim 32 --backbone distilbert --save-dir results/distilbert/ --init-mode static --beta-i 0.0 --beta-g 1.0 --beta-s 1.0 --imask-dropout 0.3 --project WIGRAPH --non-linearity gelu --num_train_epochs 10 --seed 42 --onlyA --max_sent_len 56 --no-save --anneal
- The aopc (after interpretation files are generated) can be measured using:
aopc.py --per_gpu_train_batch_size 32 --per_gpu_eval_batch_size 64 --task-name sst2 --learning_rate 1e-05 --factor 1000.0 --mask-hidden-dim 32 --backbone distilbert --save-dir results/distilbert/ --init-mode static --beta-i 0.0 --beta-g 1.0 --beta-s 1.0 --imask-dropout 0.3 --project WIGRAPH --non-linearity gelu --num_train_epochs 10 --seed 42 --onlyA --max_sent_len 56 --no-save --anneal
- The hyperparameters for both the above are the same as the training hyperparameters used for training the model.