ConStruM: A Structure-Guided LLM Framework for Context-Aware Schema Matching

TL;DR

This paper tackles the challenge of schema matching when column semantics require evidence beyond the column itself. It introduces ConStruM, a structure-guided add-on that builds a compact, query-specific context pack from two reusable structures: a context tree for multi-level schema neighborhoods and a global similarity hypergraph for confusable groups. Offline preprocessing constructs these structures, while online instantiation assembles a tailored context pack appended to the final LLM decision prompt, enabling better grounding under a fixed budget. Empirical results show that ConStruM achieves substantial gains on the context-stress HRS-B benchmark and competitive performance on MIMIC-2-OMOP, with ablations confirming the contributions of both the hierarchical context and the contrastive grouping cues. Overall, the approach demonstrates the practical value of budgeted, structure-guided evidence packing for robust schema matching in data integration tasks.

Abstract

Column matching is a central task in reconciling schemas for data integration. Column names and descriptions are valuable for this task. LLMs can leverage such natural-language schema metadata. However, in many datasets, correct matching requires additional evidence beyond the column itself. Because it is impractical to provide an LLM with the entire schema metadata needed to capture this evidence, the core challenge becomes to select and organize the most useful contextual information. We present ConStruM, a structure-guided framework for budgeted evidence packing in schema matching. ConStruM constructs a lightweight, reusable structure in which, at query time, it assembles a small context pack emphasizing the most discriminative evidence. ConStruM is designed as an add-on: given a shortlist of candidate targets produced by an upstream matcher, it augments the matcher's final LLM prompt with structured, query-specific evidence so that the final selection is better grounded. For this purpose, we develop a context tree for budgeted multi-level context retrieval and a global similarity hypergraph that surfaces groups of highly similar columns (on both the source and target sides), summarized via group-aware differentiation cues computed online or precomputed offline. Experiments on real datasets show that ConStruM improves matching by providing and organizing the right contextual evidence.

Figures (4)

• Figure 1: A context-critical schema-matching example from MIMIC-III: CHARTTIME is ambiguous between “time observed” and “time recorded/entered” targets; the sibling column STORETIME provides the clue to choose the right meaning.
• Figure 2: Three approaches to LLM-based schema matching. (a) Local: the LLM reasons over an isolated source–target column pair, with little or no additional context. (b) Global: the LLM is provided with the full source and target schemas in a single prompt. (c) Structure-guided (our approach): schema-derived structure is used to assemble, for each query column, a compact, query-specific set of contextual evidence that is supplied to the LLM for matching.
• Figure 3: Multi-level context retrieval with a context tree (depth varies by dataset; we show three levels for readability). For a query column, ConStruM retrieves a column-to-root lineage of increasingly coarse summaries, providing both local cues and broader schema context under a fixed prompt budget.
• Figure 4: Three-stage use of similarity groups (schematic; shown for target candidates). Hyperedges represent confusable groups (sets) of columns. (a) An upstream method selects a small shortlist (e.g., top-$k$ by embedding cosine similarity). (b) ConStruM optionally expands it by adding group members of strong candidates, then generates group-wise contrast cues for confusable groups within the candidate set. (c) The final LLM decision consumes the shortlist, context, and cues to select the Top-1 target. (We apply the same grouping-and-contrast idea to the source/query side to clarify the query column when it has confusable siblings/near-duplicates.)