Rethinking Language Bias in Referring Image Segmentation via Linguistically Enriched Data Reconstruction and Semantic Alignment

Referring Image Segmentation (RIS) enables fine-grained interaction between vision and language by segmenting target objects based on natural language expressions. However, existing RIS models are fundamentally constrained by the linguistic bias of manually annotated benchmarks, whose captions exhibit limited lexical, syntactic, and contextual diversity. As a result, these methods often overfit to dataset-specific linguistic patterns and struggle with real-world expressions that deviate from these distributions. To address this issue, we propose a linguistically enriched RIS framework that improves robustness under diverse linguistic expressions. On the data side, we introduce a multi-stage MLLM-driven caption reconstruction pipeline that automatically expands linguistic richness of existing RIS datasets through instance-specific generation, relational context integration, location-aware semantic verification, and linguistic diversification enhancement. On the model side, we design two complementary enhancement strategies: (i) a Decoupled Text-Visual Interaction (DTVI) module that refines fine-grained cross-modal alignment through component-level textual decomposition, and (ii) a Semantic-Aware Mosaic Augmentation (SAMA) strategy that leverages dual-caption semantic cues to mine informative negative contexts and improves target-distractor discrimination under visually confusing contexts. Extensive experiments on three public RIS benchmarks and three representative architectures demonstrate that our framework achieves consistent performance improvement on both original and reconstructed datasets.