Synthetic tabular data can replicate the distribution of real datasets but often break the semantic constraints that define valid rows. This reveals a fundamental limitation of current tabular generators: they focus on distributional fidelity while ignoring weak semantic priors encoded in schemas and textual descriptions.
To address this, we introduce a semantics‑consistent tabular diffusion framework (denoted as \ours) that generates high‑fidelity data under weak semantic priors. \ours first extracts two types of priors— intra‑column semantics and inter‑column symbolic rules—using a large language model on metadata and validates them on the real training split. These priors serve as generation conditions rather than post‑hoc filters.
The framework maps heterogeneous column values, column identities, and semantic priors into a unified semantic space, then performs column‑wise forward corruption followed by prior‑conditioned reverse denoising. This preserves marginal distributions while enforcing rule‑consistent cross‑column dependencies, achieving both distributional fidelity and semantic consistency.
Extensive experiments on six real‑world tabular benchmarks show that \ours consistently outperforms representative VAE, GAN, LLM, and diffusion baselines in distributional similarity, semantic consistency, and downstream task utility. Additional analysis demonstrates that \ours remains robust even when some semantic priors are unavailable.
Review