Syntax-Aware Dependency Parsing for Dual-Origin Noisy Correspondence in Text-Based Person Search

Text-based person search (TBPS) aims to retrieve a target individual from an image gallery based on a natural language description. In practice, this cross-modal retrieval task is often affected by noisy correspondence (NC) arising from annotation inaccuracies and low-quality visual observations. Such noise introduces two key challenges: (1) textual ambiguity within the description, including unclear attribute-Adjective associations; and (2) semantic confusion across modalities, where visually similar individuals are described with overlapping language, leading to feature entanglement and retrieval ambiguity. To tackle these challenges, we propose a Dual-level Adaptive Robust Network (DARN) that enhances both syntactic and geometric robustness during training. At the textual level, a syntax-aware text (SAT) purification module leverages dependency parsing and relevance-adaptive token selection to clarify modifier scopes and suppress noisy or low-relevance words. At the alignment level, a geometry-robust aligner (GRA) applies similarity variance regularization and identity-balanced contrastive learning to prevent feature collapse among visually similar identities. Furthermore, a noise-coupled loss dynamically adjusts the contribution of each sample according to a noise coefficient derived from syntactic uncertainty and semantic inconsistency, enabling adaptive optimization under varying noise intensities. Extensive experiments on CUHK-PEDES, ICFG-PEDES, and RSTPReid under different noise injection levels (0%, 20%, and 50%) demonstrate that DARN consistently improves retrieval accuracy and alignment stability over CLIP-based baselines. Our method consistently delivers strong performance across all three datasets, achieving 76.45% R@1 and 52.00% mINP on CUHK-PEDES, and maintaining high robustness even in the presence of synthetic correspondence noise.