GPT-RE: In-context Learning for Relation Extraction using Large Language Models

This is the official repository for the paper "GPT-RE: In-context Learning for Relation Extraction using Large Language Models" (EMNLP 2023).

Updates

[New] The GPT-RE with Fine-Tuned representations (GPT-RE_FT) method is now available! The code has been reproduced and organized in our follow-up work. Please visit:

GPT-RE-FT Repository

This repository contains the complete implementation including fine-tuned representation methods that were not included in the original release.

Supported Features

This repository supports the following methods from the paper:

• GPT-Random: Random demonstration selection
• GPT-SimCSE: SimCSE-based similar demonstration retrieval
• Reasoning Logic: Task-aware reasoning in demonstrations
• Entity-aware Similarity: Entity information enhanced retrieval

Usage

bash run_relation_ace.sh

Configuration Options

In the example script file run_relation_ace.sh, the following options are available:

Basic Settings

Option	Description	Example
--task	Name of the task	semeval, ace05, scierc
--model	GPT model name	text-davinci-003
--seed	Random seed	42

Data Settings

Option	Description
--num_test	Number of test examples (should be smaller than test dataset size)
--example_dataset	File path for demonstration examples
--test_dataset	Path to test data (use test.json for full test set)

Demonstration Settings

Option	Description
--fixed_example	1: Fixed demonstrations; 0: Re-retrieve for each test (use 0 for kNN)
--fixed_test	1: Fixed test dataset (can be ignored, use --test_dataset instead)
--num_per_rel	Examples per relation type for demonstrations (use 0 for kNN)
--num_na	NA examples for demonstrations (use 0 for w/o NA and kNN setups)
--num_run	Keep 1

Method Settings

Option	Description
--random_label	1: Use random labels in demonstrations
--reasoning	1: Add reasoning to demonstrations
--use_knn	1: Use kNN for demonstration retrieval
--k	Top-k for kNN retrieval
--reverse	1: Reverse demonstration order (default 0: more similar at top)
--entity_info	Entity-aware sentence similarity (from our paper)

Advanced Options (can be ignored)

Option	Description
--var	Keep 0
--verbalize	Keep 0
--structure	Structured prompt trial, keep default
--use_ft	Fine-tuned representation - see GPT-RE-FT for full support
--self_error	Keep 0
--use_dev, --store_error_reason, --discriminator	Keep 0

Datasets

The repository includes preprocessed data for:

• SemEval 2010 Task 8
• ACE05
• SciERC

Citation

If you find this work helpful, please cite our paper:

@inproceedings{wan2023gpt,
  title={GPT-RE: In-context Learning for Relation Extraction using Large Language Models},
  author={Wan, Zhen and Cheng, Fei and Mao, Zhuoyuan and Liu, Qianying and Song, Haiyue and Li, Jiwei and Kurohashi, Sadao},
  booktitle={Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing},
  year={2023}
}

License

Please refer to the LICENSE file for details.