An Empirical Comparison of Instance Attribution Methods for NLP

This repository is the implementation for the paper "An Empirical Comparison of Instance Attribution Methods for NLP"

Instruction

This project uses Allennlp + Pytorch.

To install the requirements, use the conda_env.yml file .

I am storing data in the Datasets/ folder (you can also use some other folder but you need to set a DATADIR environment variable ). For any dataset (for example, SST), the data should be available in following dirs -

1. Datasets/{dataset_name}/data/{train,dev,test.jsonl} (or {DATADIR}/{dataset_name}/data/{train,dev,test.jsonl})
2. Each line in jsonl file is of form

{
    "idx" : str,
    "document": str,
    "query": Optional[str],
    "label": str
}

For NLI, we could generate dataset in above format by setting for each example, the document keys as premise [SEP] hypothesis .

The code is stored in the influence_info dir. The structure is as follows :

1. dataset_readers - Contains dataset reader (see Allennlp) for use in our models. The reader for text data is base_reader.py . Code should be self explanatory.

2. models - Contains models we want to train in classifiers subfolder. For text data, we want to use the transformer_text_classifier.py . This basically uses a BERT model with a linear classifier on top. We use the pretrained transformer embedder from Allennlp.

3. training_config - Contains the jsonnet configuration files for training models. For text data, we want to use the transformer_text_classifier.jsonnet. Some configurations can be set by using environment variables. See std.extVar statements in the file. Should be self explanatory.

4. Most of my code is run using bash scripts in the commands folder.

   1. For training a model, the file to use is train.sh (which calls base_train.sh after setting some paths). It takes a few environment variables,

   DATADIR=Datasets \
   OUTPUT_DIR=outputs \
   CUDA_DEVICE=0 \
   DATASET_NAME=SST \
   CLASSIFIER=transformer_text_classifier \
   EPOCHS=10 \
   BSIZE=8 \
   EXP_NAME=<your-experiment-name> \
   bash influence_info/commands/train.sh
   

   This will store your model in {OUTPUT_DIR}/{DATASET_NAME}/{EXP_NAME}/{CLASSIFIER} .

   Note this code also makes predictions on your train,dev,test files and store them the same output folder as predictions.{train,dev,test}.jsonl files (see base_predict.sh)

5. One you have a trained model, you can run all attribution methods (and all their subtypes) on it by just using the influence_all.sh file in commands folder.

   BSIZE=8 \
   EXP_NAME=<your-experiment-name> \
   DATASET_NAME=SST \
   DATADIR=Datasets \
   CUDA_DEVICE=0 \
   OUTPUT_DIR=outputs \
   CLASSIFIER=transformer_text_classifier \
   bash influence_info/commands/influence_all.sh
   

   For each influence method type, the results are stored in folder {OUTPUT_DIR}/{DATASET_NAME}/{EXP_NAME}/{CLASSIFIER}\{name of the influencer}\{influencer configuration}\

   This contains 3 file -

   1. influence_values.npy - This is a numpy matrix that can be loaded using np.load. It is a matrix of size (Validation set size, Training set size, number of classes) .
   2. training_idx.json - Maps the index of each element in the second dimension above to idx field in training data file.
   3. validation_idx.json - Maps the index of each element in the first dimension above to idx field in validation data file.

6. To run the influence function (complete), use

   BSIZE=1 \
   EXP_NAME=<your-experiment-name> \
   DATASET_NAME=SST \
   DATADIR=Datasets \
   CUDA_DEVICE=0 \
   OUTPUT_DIR=outputs \
   CLASSIFIER=transformer_text_classifier \
   INFLUENCER=influence_function \
   PARAM_REGEX=".*" \
   bash influence_info/commands/influence.sh