KAB: A knowledge-aligned benchmark for reproducible evaluation of distantly supervised relation extraction

Structured knowledge is crucial for enhancing relation extraction (RE), yet existing datasets lack standardized integration with knowledge graphs, limiting fair evaluation and comparative analysis of knowledge-aware methods. To address this gap, we propose KAB, a benchmark for knowledge-enhanced distantly supervised relation extraction. KAB is constructed through a unified two-stage framework: first, hybrid entity linking and structural context retrieval from external knowledge bases, followed by dataset-level refinement and alignment. Each instance is enriched with graph-based semantic neighborhoods from Wikidata or Freebase, enabling RE models to leverage multi-hop and relational context without altering the original task formats. Experimental results show consistent performance improvements across models on the KAB dataset, with significant gains in high-coverage knowledge graph scenarios. The benchmark also highlights how model performance varies with the structure and density of the knowledge graph, emphasizing the substantial impact of knowledge integration on relation extraction effectiveness.