Knowledge-Aware Graph-Enhanced Transformer for Semantic Retrieval

Neural information retrieval has transformed search systems through powerful contextual embeddings, yet struggles persist with vocabulary mismatch and lack of explicit relational knowledge. A knowledge-aware framework combines transformer-based semantic encoding with graph-structured reasoning to significantly improve document ranking accuracy. The approach automatically constructs a corpus-level knowledge graph from entity relationships, generates dense embeddings via bi-encoders with synonym expansion, and employs graph convolutional networks for multi-hop relational reasoning. Contrastive learning then aligns relevant query-document pairs while enhancing robustness. Evaluated on the MS MARCO benchmark, the method consistently outperforms lexical and dense retrieval baselines, achieving substantial gains in NDCG@10, MRR@10, and Recall@1000. These results demonstrate that integrating structured knowledge with neural representations enhances both retrieval effectiveness and interpretability, paving the way for more robust large-scale information retrieval systems.