A Variational Approach for Mitigating Entity Bias in Relation Extraction

2025-06-13

AI Summary

Relation extraction (RE) models suffer from entity bias, over-relying on superficial entity surface forms and exhibiting poor generalization. To address this, we propose the first application of the Variational Information Bottleneck (VIB) to RE, explicitly disentangling entity-specific information in the latent space while preserving relation-discriminative features. We further integrate adversarial entity-agnostic representation learning to enhance robustness against entity perturbations. Our approach is theoretically grounded, interpretable, and inherently adaptable across domains. Extensive experiments demonstrate state-of-the-art performance on three major RE benchmarks—general-domain, financial, and biomedical—achieving superior in-domain accuracy and exceptional cross-domain generalization under type-constrained substitutions (e.g., entity swapping and domain transfer), significantly outperforming existing methods.

Technology Category

Reasoning under Uncertainty: Relational Probabilistic ModelsNatural Language Processing: Information ExtractionKnowledge Representation and Reasoning: Diagnosis and Abductive Reasoning

Application Category

Search and Retrieval-Augmented AI: Vertical and domain-specific searchGraph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphsSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Mitigating entity bias is a critical challenge in Relation Extraction (RE), where models often rely excessively on entities, resulting in poor generalization. This paper presents a novel approach to address this issue by adapting a Variational Information Bottleneck (VIB) framework. Our method compresses entity-specific information while preserving task-relevant features. It achieves state-of-the-art performance on relation extraction datasets across general, financial, and biomedical domains, in both indomain (original test sets) and out-of-domain (modified test sets with type-constrained entity replacements) settings. Our approach offers a robust, interpretable, and theoretically grounded methodology.

Problem

• Mitigating entity bias in Relation Extraction models
• Compressing entity info while keeping task features
• Improving generalization across multiple domain datasets

Innovation

• Uses Variational Information Bottleneck framework
• Compresses entity-specific information
• Preserves task-relevant features