Grammar-Constrained Decoding for Structured NLP Tasks without Finetuning

---

🌟 New Implementation Compatible with HuggingFace Transformers

We provide an implementation of GCD that is compatible with the popular Transformers library!

This new package, Transformers-CFG, extends the capabilities of our Grammar-Constrained Decoding (GCD) approach by integrating seamlessly with the Transformers library. It offers:

• Easy Integration: Quickly combine the power of GCD with any model listed in the transformers library with just few lines of code!
• Enhanced Performance: Leverage the GCD technique for more efficient and accurate generation.
• Friendly Interface: Implemented with the EBNF grammar interface, making it accessible for both beginners and experts.

Get started with Transformers-CFG here.

---

1. The Overview of GCD

2. Environment Setup

With the repository cloned, we recommend creating a new conda virtual environment:

conda create -n GCD python=3.9
conda activate GCD

Install the required packages:

pip install -r requirements.txt

Experiments

• Download datasets, grammars and models
• Build task-specific grammars
• Windows-specific setting
• Running the experiments

Citation

This repository contains the code for the models and experiments in Grammar-Constrained Decoding for Structured NLP Tasks without Finetuning

@inproceedings{geng-etal-2023-grammar,
	title        = {Grammar-Constrained Decoding for Structured {NLP} Tasks without Finetuning},
	author       = {Geng, Saibo  and Josifoski, Martin  and Peyrard, Maxime  and West, Robert},
	year         = 2023,
	month        = dec,
	booktitle    = {Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing},
	publisher    = {Association for Computational Linguistics},
	address      = {Singapore},
	url          = {https://aclanthology.org/2023.emnlp-main.674},
	editor       = {Bouamor, Houda  and Pino, Juan  and Bali, Kalika}
}

Please consider citing our work, if you found the provided resources useful.

License

This project is licensed under the terms of the MIT license.