Understanding Dataset Difficulty with V-Usable Information (ICML 2022, outstanding paper)

Link to Paper: https://arxiv.org/abs/2110.08420

The high-level idea is that dataset difficulty corresponds to the lack of V-usable information, a generalization of Shannon Information proposed by Xu et al. (2019). V refers to a specific function family (e.g., BERT).

Quick Start

Here's how to use this package to estimate the BERT-usable information in the SNLI dataset:

1. Install the packages in requirements.txt: pip install -r requirements.txt

2. In augment.py, instantiate the transformations and call .transform(). The base class will download the data from HuggingFace. This will create CSV files with two columns: 'sentence1' (containing the input) and 'label' (containing the label). SNLIStandardTransformation is a pass-through transformation that produces snli_train_std.csv and snli_test_std.csv; it does not change the input text in any way. SNLINullTransformation replaces the input with the null variable (i.e., empty string) to produce snli_train_null.csv and snli_test_null.csv

   import augment
   SNLIStandardTransformation('./data').transform()  
   SNLINullTransformation('./data').transform()
   

3. Use run_glue_no_trainer.py to train two BERT models: one on snli_train_std.csv; one on snli_train_null.csv. You can run the commands in finetune.sh (with bert-base-cased as the value for $MODEL). Change the values for $PROJ_DIR and $MODEL_DIR as needed.

   python run_glue_no_trainer.py \
   --model_name_or_path bert-base-cased \
   --tokenizer_name bert-base-cased \
   --train_file ./data/snli_train_std.csv \
   --validation_file ./data/snli_train_std.csv \
   --per_device_train_batch_size 32 \
   --per_device_eval_batch_size 32 \
   --num_train_epochs 2 \
   --seed 1 \
   --output_dir $MODEL_DIR/finetuned/bert-base-cased_snli_std2
   
   
   python run_glue_no_trainer.py \
   --model_name_or_path bert-base-cased \
   --tokenizer_name bert-base-cased \
   --train_file ./data/snli_train_null.csv \
   --validation_file ./data/snli_train_std.csv \
   --per_device_train_batch_size 32 \
   --per_device_eval_batch_size 32 \
   --num_train_epochs 1 \
   --seed 1 \
   --output_dir $MODEL_DIR/finetuned/bert-base-cased_snli_null
   

4. Then estimate the V-usable information using v_info.py. This will write a CSV file of the pointwise V-information (PVI) values -- for every example in the SNLI test data -- which you can average to estimate the V-usable information.

   from v_info import v_info
   
   v_info(
     f"./data/snli_test_std.csv",
     f"{MODEL_DIR}/finetuned/bert-base-cased_snli_std2",
     f"./data/snli_test_null.csv", 
     f"{MODEL_DIR/finetuned/bert-base-cased_snli_null",
     'bert-base-cased',
     out_fn=f"PVI/bert-base-cased_std2_test.csv"
   )
   

Use Cases

Our framework can be used for multiple purposes:

1. Compare different models by replacing bert-base-cased with other models (you can specify a model architecture on Huggingface or one of your own).

2. Compare different datasets by writing transformations for them in augment.py and specifying those files in steps 3-4.

3. Compare the usefulness of different attributes of the input. For example, to understand the importance of word order in SNLI, use augment.SNLIShuffleTransformation to create transformed SNLI datasets snli_train_shuffled.csv and snli_test_shuffled.csv. Then repeat steps 3-4 with the new files in place of snli_train_std.csv and snli_test_std.csv.

4. Compare the difficulty of individual examples in the dataset by looking at the PVI values in the file generated in step 5.

5. Compare the difficulty of subsets of the data by averaging the PVI values of the examples that belong to that subset, which you can do with simple pandas operations.

Reported Results

All the results reported in the paper can be found in the PVI directory. Each file is named with the format {model}_{dataset}_{transformation}_{train/test}.csv, though there are some files where the transformation and train/test split were swapped (or omitted, if there was no train/test split).

In addition to the columns present in the original dataset, each results CSV has the following columns:

• H_yb: the pointwise label entropy
• H_yx: the pointwise conditional entropy
• predicted_label: the label predicted by the model for that example
• correct_yx: whether the predicted label is correct or not
• PVI: the pointwise V-usable information (i.e., H_yb - H_yx)

You can average correct_yx to get the model accuracy and average PVI to estimate V-usable information with respect to the model V.

The data directory contains the MultiNLI, CoLA, and DWMW17 datasets with various transformations, such as shuffled word order. SNLI is not included because it is too large, though it can easily be downloaded from Huggingface. The artefacts directory contains the token-level annotation artefacts that were discovered in each dataset (see section 4.3 of the paper for details).

Citation

@InProceedings{pmlr-v162-ethayarajh22a,
  title = 	 {Understanding Dataset Difficulty with $\mathcal{V}$-Usable Information},
  author =       {Ethayarajh, Kawin and Choi, Yejin and Swayamdipta, Swabha},
  booktitle = 	 {Proceedings of the 39th International Conference on Machine Learning},
  pages = 	 {5988--6008},
  year = 	 {2022},
  editor = 	 {Chaudhuri, Kamalika and Jegelka, Stefanie and Song, Le and Szepesvari, Csaba and Niu, Gang and Sabato, Sivan},
  volume = 	 {162},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {17--23 Jul},
  publisher =    {PMLR},
  pdf = 	 {https://proceedings.mlr.press/v162/ethayarajh22a/ethayarajh22a.pdf},
  url = 	 {https://proceedings.mlr.press/v162/ethayarajh22a.html}
}