Renaming and Retrieval: Semantic Sensitivity in Join Discovery on a YADL-Augmented Data Lake

Table retrieval in data lakes often assumes that table and column names are stable and useful. This paper tests that assumption using two semantic variants of the YADL binary data lake [1]. The first variant, Join-Preserving, renames predicates but keeps all rows unchanged. The second variant, Challenge, renames predicates and reduces row overlap to about 22.5% of the original table. Across five retrieval methods, value-overlap methods are robust to renaming and only mildly affected by weaker overlap. In contrast, starmie-schema is sensitive to semantic renaming, but it does not reliably identify tables that are actually joinable. We therefore test weighted starmie, a hybrid method that combines schema similarity with a small sample of 32 column values. This method improves real top-1 containment by about 230× over schema-only retrieval while keeping strong ranking confidence. These results suggest that semantic names are useful, but they should be combined with value evidence when the goal is join discovery.