Accepted to DASFAA 2026 (full paper)

Abstract

Overview

CRAG is a relation-aware framework for Document-level Relation Extraction (DocRE) that aligns the contextual representations of entity pairs with the semantic expressions of relation labels.

Unlike traditional DocRE approaches that rely purely on contextual attention from pre-trained language models (PLMs), CRAG explicitly incorporates relation semantics through a prompt graph constructed from evidence phrases in training documents.

This approach enables the model to reason about relations using both contextual signals and structured relation semantics.

Problem

Most existing DocRE models:

• Rely heavily on contextual attention mechanisms
• Do not explicitly model the semantics of relation labels
• Often attend to irrelevant tokens unrelated to the true relation

As a result, entity-pair representations may fail to align with the actual linguistic expressions that indicate relations.

Key Idea

CRAG addresses this issue by aligning entity pair representations with relation semantics using a prompt graph. The key idea is to represent relation labels through prototype embeddings derived from evidence phrases.

These prototypes capture diverse linguistic expressions that describe relations in training documents. We construct two types of prototypes:

Interior Prototypes

• Represent core semantics of a relation
• Derived from evidence phrases that strongly express the relation

Border Prototypes

• Capture semantic boundary expressions shared between related relations