Relation Extraction with Instance-Adapted Predicate Descriptions.

Area of Science:

• Natural Language Processing
• Information Extraction
• Machine Learning

Background:

• Relation Extraction (RE) is crucial for knowledge discovery and question answering.
• Smaller encoder models are preferred for RE, despite decoder-only models excelling in generative tasks.
• Existing methods often use fixed linear layers for predicate representations.

Purpose of the Study:

• To improve relation extraction performance using a novel dual-encoder architecture.
• To develop a method that computes instance-specific predicate representations.
• To enhance fine-tuning of smaller encoder models for RE tasks.

Main Methods:

• A novel dual-encoder architecture with joint contrastive and cross-entropy loss was developed.
• A second encoder computes instance-specific predicate representations by incorporating entity spans.
• Experiments were conducted on two biomedical and two general domain RE datasets.

Main Results:

• The proposed approach achieved F1 score improvements of 1-2% over state-of-the-art methods.
• The dual-encoder architecture demonstrated superior performance on both biomedical and general datasets.
• Ablation studies confirmed the effectiveness of the architectural components.

Conclusions:

• The novel dual-encoder architecture offers a simple yet effective approach to relation extraction.
• Instance-specific predicate representations significantly enhance RE performance.
• This method provides a valuable advancement for information extraction tasks.