Enhancing Targeted Minority Class Prediction in Sentence-Level Relation Extraction.

Area of Science:

• Natural Language Processing
• Machine Learning

Background:

• Sentence-level relation extraction (RE) faces challenges with imbalanced data, where 80% of instances are negative (no relation).
• Minority classes (MC) within positive labels are often underrepresented and can suffer from label noise, leading to poor model performance.

Purpose of the Study:

• To address the challenges of label noise and low source availability in RE.
• To improve the performance of models on minority classes, which are inadequately addressed by existing methods focusing on micro F1 scores.

Main Methods:

• Introduction of a novel Minority Class Attention Module (MCAM) designed to handle imbalanced data in RE.
• Development of effective data augmentation methods specifically tailored for RE tasks.
• MCAM identifies reliable MC instances using confidence scores for augmentation and integrates MC information during model training.

Main Results:

• The proposed methods achieve state-of-the-art F1 scores on the TACRED dataset.
• Significant enhancement in the F1 scores for minority classes, demonstrating the effectiveness in handling underrepresented data.

Conclusions:

• The developed MCAM and augmentation strategies effectively tackle mis-classification errors for minority classes in RE.
• The approach offers a robust solution for improving the overall accuracy and fairness of relation extraction models on imbalanced datasets.