Semi-Supervised Relation Extraction Informed by Area Under the Margin Ranking and Large Language Models

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Relation extraction is crucial for knowledge discovery and graph construction.
• Acquiring labeled data for relation extraction is challenging, especially in low-resource scenarios.
• Semi-supervised methods like self-training offer solutions but can suffer from noisy pseudo-labels.

Purpose of the Study:

• To address the limitations of traditional self-training in low-resource relation extraction.
• To introduce a novel model, RE-AUM-LLM, for generating high-quality pseudo-labels.
• To improve the performance of relation extraction models with limited labeled data.

Main Methods:

• Implemented a self-training framework combined with Area Under the Margin (AUM).
• Integrated Large Language Models (LLMs), specifically Llama 3.1, for pseudo-label generation.
• Evaluated the RE-AUM-LLM model on two benchmark datasets for low-resource relation extraction.

Main Results:

• The RE-AUM-LLM model achieved state-of-the-art performance on benchmark datasets.
• The proposed approach significantly improved relation extraction accuracy in low-resource settings.
• High-quality pseudo-labels generated by RE-AUM-LLM mitigated noise issues in self-training.

Conclusions:

• RE-AUM-LLM offers a powerful solution for low-resource relation extraction.
• Combining AUM and LLMs enhances pseudo-label quality and model performance.
• The approach demonstrates significant advancements in leveraging limited data for relation extraction tasks.