Skip to content

Latest commit

 

History

History
59 lines (40 loc) · 3.33 KB

README.md

File metadata and controls

59 lines (40 loc) · 3.33 KB

Embedding Recycling

Preprint on ArXiv: Embedding Recycling for Language Models

Datasets

To access our datasets for text classification, please go to the folder titled text_classification.

To access our datasets for named-entity recognition (NER), please go to the folder titled ner.

To access our datasets for question answering (QA), please access the TriviaQA and SQuAD datasets on HuggingFace at the following links:

Setup Environment

Run the following commands to setup a conda environment:

conda create --name embedding_recycling --file requirements.txt
conda activate embedding_recycling

Experiment Replication

Standard Embedding Recycling

To replicate our results for standard embedding recycling, use the conda environment listed above and run the following scripts for each dataset group:

  • For text classification, use the TextClassificationScripts/GeneralLinearClassifier_PaperResults.py script
  • For NER, use the NER_Scripts/General_NER_Classifier_PaperResults.py script
  • For QA, use the QA_Scripts/GeneralQuestionAnswering_PaperResults.py script to replicate the TriviaQA results. For TriviaQA, use the QA_Scripts/PrepareTriviaQADataset.py for preprocessing the TriviaQA dataset using the selected model. For SQuAD, please use the "run_squad.py" script included on the HuggingFace Transformers repository, which we also include in our main directory.

The hyperparameters for replicating each experiment are included in the HyperparameterSelection folder.

Adapter-Based Embedding Recycling

To replicate our results for adapter-based embedding recycling, use the conda environment listed above and run the following scripts for each dataset group:

  • For text classification, use the TextClassificationScripts/Adapters_PaperResults.py script
  • For NER, use the NER_Scripts/Adapters_NER_PaperResults.py script
  • For QA, use the QA_Scripts/Adapters_QA_PaperResults.py script to replicate the TriviaQA results. For TriviaQA, use the QA_Scripts/PrepareTriviaQADataset.py for preprocessing the TriviaQA dataset using the selected model. For SQuAD, please use the "run_squad.py" script included on the HuggingFace Transformers repository, which we also include in our main directory.

Citing

@misc{https://doi.org/10.48550/arxiv.2207.04993,
  doi = {10.48550/ARXIV.2207.04993},
  url = {https://arxiv.org/abs/2207.04993},
  author = {Saad-Falcon, Jon and Singh, Amanpreet and Soldaini, Luca and D'Arcy, Mike and Cohan, Arman and Downey, Doug},
  keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Embedding Recycling for Language Models},
  publisher = {arXiv},
  year = {2022},
  copyright = {Creative Commons Attribution 4.0 International}
}

EmbeddingRecycling is an open-source project developed by the Allen Institute for Artificial Intelligence (AI2). AI2 is a non-profit institute with the mission to contribute to humanity through high-impact AI research and engineering.