Syntax-Enhanced Contextual Representation Learning for Complex Relation Extraction via Attention-Guided Graph Network

Abstract

Relation Extraction (RE) aims to identify semantic relations between entity pairs in text. Pre-trained language models (PLMs) such as BERT provide strong contextual representations, but they may struggle with long-distance dependencies and irrelevant context because they do not explicitly encode syntactic structure. Dependency graphs can inject structural information, yet many graph-based RE models rely on GCN-style aggregation, which is sensitive to noisy parses and does not always integrate semantic and structural features effectively. In this paper, we propose SE-AGGN, a syntax-enhanced framework that combines PLM semantics with dependency structures in a noise-aware and adaptive manner. We employ a Graph Attention Network (GAT) over dependency graphs to softly weight neighboring nodes and reduce error propagation from unreliable edges. We further introduce a semantic-structure fusion attention module to adaptively integrate contextual representations from PLMs with structural features from the graph encoder. Experiments on SemEval-2010 Task 8 and TACRED show that SE-AGGN consistently outperforms the reproduced and commonly used baselines considered in this work, achieving 90.1 Macro-F1 on SemEval using the official SemEval scorer and 71.2 Micro-F1 on TACRED under the standard TACRED evaluation setting. Additional analyses further demonstrate improved robustness under limited supervision and degraded dependency structures.