GaV: Guess and Verification of Column Semantics

Abstract

Understanding column semantics in tabular datasets is a fundamental yet costly task that often relies on human interpretation of ambiguous attributes. We propose GaV (Guess and Verification), a multi-agent architecture that automates this process through an iterative hypothesis generation-and-verification paradigm. A Hypothesis Generator formulates candidate semantic statements for each column with respect to a user-specified semantic aspect (from a configurable aspect inventory), while a Verifier empirically validates them against the data. This mechanism grounds large language model (LLM) reasoning in empirical evidence, yielding verifiable and interpretable semantic profiles. We formalize the Aspect-Based Column Understanding (ABCU) task and introduce a benchmark comprising 46 datasets. Experimental results show that GaV attains up to 88.2% accuracy, effectively balancing accuracy and cost. By coupling LLM-agents with data-grounded validation, GaV establishes a scalable and transparent framework for semantic data understanding.