ColBERT-AW: Enhancing Late Interaction Retrieval With Attribute-Aware Query Token Weighting
Gi-taek An, Kyung Soon Lee · IEEE Access · 2026
In product search, fine-grained cues within user queries—such as attribute terms and constraint expressions—play a crucial role in determining ranking quality. However, existing dense and late-interaction retrieval models, including ColBERT, assign uniform importance to all query tokens, thereby failing to capture attribute sensitivity. To overcome this limitation, we propose ColBERT-AW (Attribute-aware Query-Token Weighting), a lightweight extension that injects token-specific scalar weights into ColBERT’s late-interaction computation. The module operates exclusively on the query side, preserving document embeddings and the index, and introduces only a minimal computational overhead through a single MLP pass. During training, attribute-token annotations generated by a large language model (LLM) guide the model to learn attribute-aware token weighting through a dual-objective loss that jointly optimizes token- and document-level supervision. Experiments on the TREC 2023 Product Search test set demonstrate consistent and significant improvements, achieving an NDCG@10 of 0.7584, a +0.0496 absolute gain (+7.0%) over the baseline ColBERT. These findings confirm that augmenting ColBERT with attribute-aware query weighting substantially enhances retrieval effectiveness while maintaining efficiency, scalability, and deployability, especially in scenarios requiring fine-grained understanding of attribute and constraint signals.