dhananjayashok / scifix Goto Github PK

SciFix is the official implementation of The student becomes the master: Outperforming GPT3 on Scientific Factual Error Correction. We release the dataset setup and code to replicate results and apply our method to a new dataset.

For more details, refer to the accompanying paper: The student becomes the master: Outperforming GPT3 on Scientific Factual Error Correction. If you have questions, please feel free to reach us at [email protected] / [email protected] or open an issue.

If you find this repository useful or use this code in your research, please cite the following paper:

Ashok, Dhananjay, et al. "The student becomes the master: Outperforming GPT3 on Scientific Factual Error Correction." The 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

@inproceedings{ashok2023student,
  title={The student becomes the master: Outperforming GPT3 on Scientific Factual Error Correction},
  author={Ashok, Dhananjay and Kulkarni, Atharva and Pham, Hai and Poczos, Barnabas},
  booktitle={The 2023 Conference on Empirical Methods in Natural Language Processing},
  year={2023}
}

Set up the Python (3.10) environment through conda using:

conda env create -f environment.yml

We take the SciFact-Open dataset and place the jsonl files under SciFix/scifact-open/data, similarly, the CovidFact jsonl is stored as SciFix/covidfact/COVIDFACT_dataset.jsonl. See data.py for parsing implementation, if you want to implement a new dataset you will be interested in the function to_common_format, if you make sure your dataset is eventually in this format and then passed in a list to get_joined_df you should not have to make any other changes to the code.

Once in the common format, we can generate the synthetic datasets using generate_full_dataset and split them using generate_splits. To change the underlying prompt model see query_models.py and dataset_creation.py. To modify the prompts used see prompting.py

We use the files finetune_as_summarization.py to generate the Seq2Seq Language Model and finetune_classification.py to train the semantic difference model

See the query_models.py file and refer to get_bart_cc() as an example of loading a trained Seq2Seq Error Correction Model, to use score-guided decoding see score_guided_decode. The main function has an example of these functions being used.

This repository is licensed under the terms of the Apache License.