About

This is the code release to support the following paper, accepted at ACL2021: Evidence-based Factual Error Correction by James Thorne and Andreas Vlachos, University of Cambridge.

Training/Testing Example

Data

Can be downloaded from this Google Drive folder. Thank you for Nicola de Cao for generating the GENRE predictions, they have been merged with the predictions from DPR using the code release from this Github Repo. The merge script is in src/error_correction/dataset_utils/filter_dpr_genre.py.

Environment

Create a new conda repository and pip install -r requirements.txt. All experiments expect src to be in the pythonpath:

export PYTHONPATH=src

License

The source code released in this repository is released under the Apache 2.0 license. Some portions of the training script are adapted from examples released by HuggingFace, which is also released under an Apache 2.0 license and modified to support the error correction task.

The annotations for the FEVER dataset are released under the Creative Commons Share Alike License v3.0. More information can be found here https://fever.ai/dataset/fever.html

Running Experiments

Masker

For all maskers, they can be used with either the gold evidence from the master file, or if pipeline_text is present in the JSON document for each instance, the predicted evidence will be used for generating masks.

LIME-based masker (black-box)

Requires a pre-trained HF model that is trained as a (claim, evidence) pair classifier. This is quite slow though, it would be best to break the input file into multiple chunks or set it running and come back a day later (esp. when generating predictions for training dataset).

To first train the FEVER task classifier that will be used to generate lime predictions:

bash scripts/finetune_fever.sh 5e-5 fever_sub genre_50_2 true

Then:

python -m error_correction.masking.lime_masker <in_file> <out_file> --model <path/to/huggingface/classifier>

Random masker

Optional argument --sample_prob=0.15, defines the number of tokens masked from the claim.

python -m error_correction.masking.random_masker <in_file> <out_file>

Masked-Language-Model masker

Required argument --top_k=3, defines the number of tokens to be masked (k tokens from MLM with highest per-token loss).

First use MLM to generate per-token mask predictions and save losses to mlm_file

python -m error_correction.masking.generate_mlm_losses <in_file> <mlm_file>

Then select the top k tokens from this file to use as an explanation

python -m error_correction.masking.lang_model_masker <mlm_file> <out_file> --top_k=3

ESIM-based masker (white-box)

In a separate folder/conda environment, follow the instructions from this GitHub repository and run the model_predictions.py file on appropriate data split. Several changes have been made to this repository, apply the patch in third_party_patches/ from the TokenMasker repository. git apply ../path/to/third_party_patches/schuster.patch. There is an additional model_predictions file for using retrieved evidence instead of gold evidence called model_predictions_pipeline.py.

Then convert the predictions with this script select the top k tokens from this file to use as an explanation

python -m error_correction.masking.schuster_convert_masker <predictions_file> <out_file>

Corrector

Fully-supervised ceiling

With gold evidence

bash scripts/finetune_supervised.sh t5-base 1e-4 true false all

With predicted evidence (from genre_dpr)

bash scripts/finetune_supervised_pipeline.sh t5-base 1e-4 true false all genre_dpr

To make predictions:

python -m error_correction.corrector.run \
    --model_name_or_path t5-base \
    --output_dir <path_to_trained_model_directory> \
    --do_predict \
    --test_file <path_to_test_file> \
    --reader supervised \
    --mutation_source true \
    --mutation_target false \
    --labels all

Masker-corrector

Expects masked instances to be stored in resources/masker/{masker_name}_{split}.jsonl for gold evidence and resources/masker/{masker_name}_{split}_{ir_method}.jsonl for evidence selected from IR system