Dense retrieval has become the dominant paradigm for large-scale table retrieval. However, existing contrastive retrievers often suffer from bag-of-words degradation: they rely on lexical overlap rather than row–column bindings. To quantify and mitigate this, we propose an adversarial benchmark and a structure-aware retrieval framework. Our benchmark uses "evidence cell" anchoring and a Homologous Structural Swapping mechanism to construct lexically equivalent but structurally invalid negative samples. Our model uses text generation as a structural regularizer for dense table representations. It decouples row and column anchors through orthogonal attention pooling, and employs dynamic masking to encourage autoregressive reconstruction. This process encourages the retrieval representations to retain table content and structure-related cues. Extensive experiments on OTT-QA, NQ-Tables, and our adversarial benchmark demonstrate the effectiveness of our method. It achieves competitive performance on conventional retrieval datasets and substantially improves structural robustness under adversarial settings.

Table-to-text generation aims to automatically produce natural language descriptions from structured or semi-structured tabular data. Unlike traditional text generation tasks, it requires models to accurately understand and represent table structures. Existing approaches typically process tables by linearizing them or converting them into graph structures. However, these methods either fail to adequately capture the table structure or rely on complex attention mechanisms, limiting their applicability. To tackle these challenges, we propose QuASAR, a question-driven self-supervised approach designed to enhance the model’s structural perception and representation capabilities. Specifically, QuASAR formulates a set of structure-related queries for self-supervised training, explicitly guiding the model to capture both local and global table structures. Additionally, we introduce two auxiliary pre-training tasks: a word-to-sentence reconstruction task and a numerical summarization task, which further enhance the fluency and factuality of the generated text. Experimental results on the ToTTo and HiTab datasets demonstrate that our approach produces higher-quality text compared to existing methods.