State of Relation Extraction using Large Language Models

Overview

This repository contains a comprehensive survey of recent advancements in Relation Extraction (RE) using Large Language Models (LLMs). The paper explores cutting-edge techniques and methodologies that are transforming the field of information extraction.

Key Contributions

The research provides an in-depth analysis of three primary approaches to relation extraction with LLMs:

1. Prompt Design

   • Explores techniques like Chain of Thought (CoT)
   • Demonstrates how carefully crafted prompts can improve model performance

2. Alignment Techniques

   • Addresses challenges in low-incidence tasks
   • Introduces innovative approaches like QA4RE and RAG4RE
   • Shows how reformulating relation extraction can unlock LLM capabilities

3. Universal Information Extraction (UIE)

   • Proposes a unified framework for information extraction tasks
   • Aims to break down silos between different information extraction approaches

Key Findings

• Significant performance improvements across multiple benchmarks
• Successful techniques for addressing LLM limitations in relation extraction
• Promising directions for future research in information extraction

Datasets Explored

• DocRED
• TACRED
• New York Times Annotated Corpus
• CoNLL04
• ACE 2005

Future Research Directions

• Improved pre-training strategies
• Multilingual dataset integration
• Document-level relation extraction
• Knowledge base-aware models
• Unified information extraction frameworks

Citation

If you use this work in your research, please cite the original paper:

@article{Sachan2023RelationExtraction,
  title={State of relation extraction using LLMs: A report},
  author={Vangmay Sachan and Yanfei Dong},
  year={2023}
}

Acknowledgements

This research was conducted as part of the Odyssey 2023/2024 program at the National University of Singapore.

Feel free to contact me for further information!