Source-linked AI summary
Correspondences as Decisions: JevNexus for Decision-Centric Schema Matching
Runze Li, Hanchen Wang, Ying Zhang, Wenjie Zhang
TL;DR
Schema matching increasingly relies on generative reranking even though the task is a bounded correspondence decision. JevNexus combines typed pairwise decisions with schema and instance evidence, invoking listwise refinement selectively; across six benchmark families it is statistically comparable to Magneto on MRR and Hits@1 while reducing mean runtime 7.750×.
Problem
Generative reranking expresses correspondence judgments indirectly and does not explicitly preserve complementary semantic, schema, and instance evidence or distinguish clear rankings from cases needing joint comparison.
Source: §1 · p. 2Method
JevNexus localizes candidates, scores each pair with typed semantic decisions, fuses schema and instance evidence, and selectively invokes listwise comparison when evidence disagrees and the ranking remains uncertain.
Source: §1 · p. 2, §3.1 · p. 3Results
Across six benchmark families, JevNexus is statistically comparable to Magneto in MRR and Hits@1 while reducing mean online runtime by 7.750×, although Magneto attains higher Recall@GT.
Source: §6 · p. 10, §7 · p. 11Takeaways & Limitations
Evidence fusion provides a consistent benefit, while selective refinement is most useful when evidence sources disagree and the resulting ranking remains uncertain.
Source: §7 · p. 11The evaluation uses one decision backbone and one listwise refinement model, and broader testing is needed to establish how the observed trade-off transfers.
Source: §6 · p. 11
Abstract
from arXiv · showhide
Schema matching increasingly uses generative language models to rerank retrieved column candidates, although the underlying task is a bounded correspondence decision. We present JevNexus, which combines typed pairwise decisions with schema/instance evidence and invokes listwise refinement only when the evidence disagrees and the fused margin is small. The evaluation covers 561 cases from six benchmark families. JevNexus obtains dataset-macro MRR and Hits@1 of 0.930 and 0.909, compared with 0.926 and 0.903 for Magneto, while reducing mean latency from 123.452 to 15.929 seconds (7.750). Paired analysis finds no statistically significant difference in either MRR or Hits@1. The gate invokes listwise refinement for only 5.665% of source columns and avoids the degradation caused by unconditional refinement. Code and experimental artifacts are available at https://github.com/RazeenLI/JevNexus.
1 INTRODUCTION
Schema matching supports integration of independently developed datasets, but generative reranking represents correspondence judgments indirectly and does not distinguish clear rankings from residual ambiguities. JevNexus coordinates pairwise decisions, complementary evidence, and selective cross-candidate comparison.
1 INTRODUCTION
- Generative matchers emit candidate names and scores autoregressively, while schema matching is fundamentally a bounded correspondence decision that can allow multiple or no matches.The generated judgments also require output parsing and represent correspondence evidence indirectly.Source: §1 · pp. 1–2
- In the figo_grade example, the token favors figo_stage while the grade attribute type favors tumor_grade, illustrating why a single surface-similarity signal may not resolve correspondence.The example motivates coordinating individual pair judgments, evidence integration, and comparison of unresolved candidates.Source: §1 · p. 1
- JevNexus scores each localized source-target pair separately, fuses semantic decisions with schema and instance evidence, and invokes listwise reranking only when evidence disagrees and fused scores are close.Otherwise, the fused ranking remains unchanged.Source: §1 · p. 2
- 123.452 to 15.929 seconds per case (7.750×): JevNexus leads dataset-macro MRR and Hits@1 across six benchmark families and is statistically comparable to the strongest generative method.The evidence-aware gate activates for only 5.665% of source columns; unconditional listwise refinement is less effective than selective refinement.Source: §1 · p. 2
- JevNexus’s contributions are a decision-centric formulation, an evidence-aware progressive architecture, and a controlled evaluation across 561 cases and multiple matcher types.The evaluation analyzes effectiveness, efficiency, component contributions, routing, dataset heterogeneity, and failure modes.Source: §1 · p. 2
2 PROBLEM FORMULATION AND MOTIVATION
The formulation separates candidate localization from correspondence judgment and final ranking. It contrasts generated reranking with separately scored pairwise evidence that can be combined with other signals.
2.1 Schema Matching and Ranked Output
- A schema correspondence is a source-target column pair representing the same underlying attribute, and a source may have zero, one, or multiple valid targets.The ranking objective places valid targets early in a complete ordering of target columns.Source: §2.1 · p. 2
2.2 Candidate-Localized Matching
- When target schemas are wide, localization limits each source column to at most k candidates, separating retrieval from discrimination among retrieved candidates.If no valid target is localized, downstream matching restricted to that set cannot recover it, so localization sets a recall ceiling.Source: §2.2 · p. 3
2.3 Generated versus Decision-Based Evidence
- Generative reranking jointly inspects the source and localized candidate set, emits identifiers and scores, then requires parsing and validation to produce a ranking.Its correspondence judgments are represented indirectly through generated tokens.Source: §2.3 · p. 3
- Decision-centric matching asks directly whether each localized source-target pair represents the same attribute and provides bounded pairwise evidence without generating or parsing textual scores.Those judgments can support candidate ranking or be combined with other matching signals.Source: §2.3 · p. 3
2.4 Decision-Centric Matching Objective
- Each q(s,t) is separately scored rather than normalized across candidates, so multiple candidates may receive strong evidence or all candidates may receive weak evidence.The values express semantic evidence about correspondence plausibility.Source: §2.4 · p. 3
- Pairwise semantic evidence does not prescribe the final matching score; a matcher may combine it with schema- or instance-level evidence before ranking valid targets.The objective is to rank valid targets early while bounding expensive inference to localized candidate sets.Source: §2.4 · p. 3
3 JEVNEXUS
JevNexus separates candidate localization, pairwise correspondence judgment, evidence integration, and residual cross-candidate comparison. It produces a complete ranking while reserving listwise refinement for ambiguous fused heads.
3.1 Architecture Overview
- JevNexus integrates typed semantic decisions and complementary schema/instance evidence, then routes ambiguous fused heads to local listwise comparison before forming the complete ranking.The architecture distinguishes candidate localization, correspondence judgment, evidence integration, and residual cross-candidate comparison.Source: §3.1 · pp. 3–4
3.2 Candidate Localization
- Candidate localization uses normalized names and sampled values for semantic retrieval, supplements it with exact-name matches, and retains at most k candidates plus an ordered tail.Retrieval determines candidate membership and initial order, not final correspondence evidence.Source: §3.2 · p. 4
3.3 Typed Correspondence Decisions
- Each localized source–target pair receives a separate bounded correspondence decision, without softmax normalization, so multiple candidates or none can receive strong evidence.The positive criterion allows different surface forms for equivalent attributes; the negative criterion covers distinct attributes despite lexical or value-level similarity.Source: §3.3 · pp. 4–5
- Batched questions retain candidate-specific contexts while encoding the source state once, avoiding one external request per pair and free-form numeric score extraction.The output contract requires one finite value in [0, 1] for each submitted candidate identifier.Source: §3.3 · pp. 4–5
3.4 Complementary Schema and Instance Evidence
JevNexus adds complementary schema and instance signals to semantic decisions, combining name and value evidence before selecting which localized pairs receive positive complementary scores.
3.4 Complementary Schema and Instance Evidence
- JevNexus computes complementary evidence with a non-generative operator combining schema- and instance-level signals, since semantic decisions alone do not exhaust available evidence.The approach follows the principle of composing heterogeneous match functions.Source: §3.4 · p. 5
3.4.1 Schema-Level Evidence.
- Character-trigram and token similarities jointly score normalized column names, capturing local spelling variation as well as separators, reordered fragments, and multiword names.For flat relational tables, the schema branch uses column-level name evidence rather than artificial hierarchy or neighborhood structure.Source: §3.4.1 · p. 5
- The instance branch compares column-value documents using TF–IDF over deterministically sampled rows, down-weighting values that occur broadly across columns.Sampling uses a fixed seed when a table exceeds its configured sample budget.Source: §3.4.2 · p. 5
3.4.3 Evidence Aggregation and Selection.
- The enabled name and instance signals are combined by a fixed weighted average with equal internal weights, then bidirectional selection applies Top-n, relative-to-best, and absolute-threshold conditions.Unselected localized pairs receive complementary score zero; this operator scores but does not add or remove candidates.Source: §3.4.3 · p. 5
3.5 Evidence Integration
JevNexus fuses semantic and complementary evidence for its default ranking, then selectively applies local listwise refinement when evidence conflicts and the leading fused candidates remain close.
3.5 Evidence Integration
- Semantic decision evidence and complementary lexical/value evidence are fused into a ranking score, not a match probability, and the resulting order is returned unless routing invokes refinement.The coefficient α controls the relative contribution of the two signals.Source: §3.5 · p. 5
3.6 Evidence-Aware Selective Refinement
- Listwise comparison is reserved for residual ambiguity rather than applied to every source column, because pairwise and complementary evidence resolve most localized rankings.The residual cases retain conflicting signals and closely balanced leading candidates.Source: §3.6 · p. 5
3.6.1 Routing Criterion. Let
- The routing criterion considers the candidates preferred by each evidence source and the margin between the two largest fused scores; disagreement alone is insufficient when fusion establishes a decisive order.A small margin indicates that leading fused alternatives remain difficult to distinguish; empty or singleton candidate sets bypass the gate.Source: §3.6.1 · pp. 5–6
3.6.2 Local Listwise Refinement.
- When routed, the refiner compares the highest-scoring fused candidates using their decision, complementary, and fused evidence, then permutes their fused score slots.Candidates outside the selected local head keep fused scores, and the stage cannot change candidate membership or the overall fused score distribution.Source: §3.6.2 · p. 6
3.7 Complete Ranking
JevNexus forms a complete target ranking by ordering localized candidates and appending the untouched target-schema tail. Its end-to-end procedure applies optional refinement only to a routed local head.
3.7 Complete Ranking
- The localized head is sorted by resulting scores, while targets outside the candidate set are appended in retained target-schema order to form the complete ranking.A ground-truth target outside the localized candidate set remains in the tail and cannot be recovered by downstream decision, integration, or refinement.Source: §3.7 · p. 6
3.8 End-to-End Procedure
- For each source column, the procedure localizes candidates, obtains typed decisions and complementary evidence, fuses scores, evaluates the routing gate, optionally refines the local Top-m, and appends the untouched tail.The complementary score matrix is computed once per table pair; the other matching operations are organized per source column, with no online parameter updates.Source: §3.8 · p. 6
3.8.1 Running Example.
- In the running example, conflicting evidence and a small fused margin route the leading candidates to refinement, which places tumor_grade ahead of figo_stage by permuting fused score slots.The remaining localized candidates retain fused scores, and targets outside the candidate set stay in target-schema order.Source: §3.8.1 · p. 6
3.9 Inference Cost
JevNexus bounds semantic matching to localized candidates, batches pairwise decisions, and computes complementary evidence for efficient fusion. Listwise requests are limited to small candidate sets when disagreement coincides with a small fused margin.
3.9 Inference Cost
- Localization limits semantic judgment to at most k candidates per source, while batching makes normal decision-request counts proportional to |S| rather than |S|k.The bounded decision interface avoids autoregressive generation of candidate identifiers, JSON syntax, or numeric score strings.Source: §3.9 · p. 6
- The complementary matcher computes its table-pair score matrix once, with constant-time lookup for each localized pair during fusion.Its scores are obtained without generative inference.Source: §3.9 · p. 6
- Listwise refinement considers at most m candidates and is invoked only when evidence disagreement coincides with a small fused margin.Unconditional refinement would issue a request for every eligible source column.Source: §3.9 · p. 7
4 EXPERIMENTAL EVALUATION
Across six benchmark families, JevNexus leads dataset-macro MRR and Hits@1 while substantially reducing mean latency relative to Magneto. The evaluation also examines component effects, routing behavior, sensitivity, and residual errors.
4.1 Experimental Setup
- The evaluation covers 561 cases across GDC-SM and five Valentine datasets, with shared candidate sets for JevNexus and Magneto to isolate downstream matching.The protocol reports dataset-macro and case-macro aggregates and assigns zero effectiveness to execution failures.Source: §4 · p. 7, §4.1 · p. 7
4.2 End-to-End Effectiveness
- JevNexus leads or ties ranking results on GDC, Magellan, and WikiData, but no method dominates every benchmark family.Unicorn leads OpenData MRR and Hits@1, while Magneto leads on ChEMBL and narrowly on TPC-DI.Source: §4.2 · p. 7, Table 2 · p. 8
- Across six benchmark families, JevNexus leads dataset-macro MRR and Hits@1, while Magneto leads Recall@GT; mean latency is 15.929 versus 123.452 seconds per case.Paired analysis finds JevNexus and Magneto statistically comparable in MRR and Hits@1.Source: Figure 5 · p. 7, §4.2 · pp. 7–8, Figure 6 · p. 8
4.3 Efficiency
- JevNexus has median and 95th-percentile latency of 14.694 and 30.029 seconds, versus 96.639 and 284.221 seconds for Magneto.Completed Magneto runs emit 5.31 million recorded output tokens; JevNexus's bounded interface emits no autoregressive score tokens.Source: Figure 6 · p. 8, §4.3 · p. 8
4.4 Component Ablation
- Complementary schema and instance evidence provides the largest consistent ablation gain; unconditional refinement harms three datasets, while selective refinement is strongest overall.Case-macro MRR is 0.944 for JevNexus and 0.928 for unconditional refinement.Source: §4.4 · p. 8, Table 3 · p. 9
4.5 Selective Routing
- The online gate opens for 766 of 13,522 source columns (5.665%), avoiding 94.335% of unconditional listwise calls.Among routed columns with a ground-truth correspondence, refinement improves ranking for 30.882% and degrades it for 9.804%.Source: §4.5 · p. 8, Table 4 · p. 9
4.6 Sensitivity Analysis
- Overall Recall@GT rises from 0.749 at α=0.000 to 0.808 at α=0.400, while increasing τ raises activation without improving Overall Recall@GT.Across refinement widths m=2, 3, and 5, Overall Hits@1 is 0.915, 0.914, and 0.911.Source: §4.6 · p. 9
4.7 Error Analysis
- At the reported operating point, 1.957% of ground-truth pairs fall outside the candidate set; on GDC, 16.602% never enter the candidate Top-20.The analysis distinguishes candidate-localization misses from downstream ordering errors and notes that calibration as a cause of the Recall@GT difference remains untested.Source: §4.7 · p. 9
5 RELATED WORK
Prior schema matchers combine heterogeneous signals, learned systems improve representations or transfer, and recent LLM systems use candidate reranking. JevNexus instead emphasizes bounded decisions, evidence fusion, and selective joint comparison.
Traditional and Composite Schema Matching
- Classical matchers combine linguistic, structural, and instance evidence (Rahm and Bernstein; Cupid; COMA), a complementarity JevNexus retains.JevNexus combines these signals with neural semantic judgments and separates evidence fusion from optional cross-candidate reasoning.Source: §5 · p. 9
Learned Schema Matching
- Learned systems improve schema matching through representations, training, or transfer, whereas JevNexus studies inference architecture after candidate localization.Examples include LSM, Unicorn, and ISResMat.Source: §5 · pp. 9–10
LLM-Based and Decision-Oriented Matching
- Unlike Magneto's generative scoring and joint reranking [17], JevNexus uses bounded pairwise decisions, fuses complementary evidence, and reserves listwise comparison for ambiguity.Its evidence-based, training-free gate invokes local refinement only when candidate-level comparison is warranted.Source: §5 · p. 10
6 DISCUSSION AND LIMITATIONS
JevNexus retains competitive ranking effectiveness while substantially reducing semantic-inference cost, with pairwise decisions as the default and selective listwise refinement reserved for unresolved ambiguities. The discussion identifies candidate coverage, score comparability, transferability, and evaluation scope as boundaries for the findings.
Effectiveness–efficiency trade-off
- 7.750× lower mean online runtime accompanies statistically comparable dataset-macro MRR and Hits@1 against the strongest tested generative method.The comparison uses a shared candidate generator.Source: §6 · p. 10
Localization and score comparability
- A ground-truth target absent from the bounded candidate set cannot be recovered by either pairwise decisions or listwise refinement.Increasing candidate width may improve coverage but also raises inference cost and exposes the matcher to more distractors.Source: §6 · p. 10
- Decision and fused scores order candidates effectively within a source query, but weaker pooled correspondence recall suggests limited comparability across queries.The passage identifies direct evaluation and calibration of cross-query scores as needed directions.Source: §6 · p. 10
Transferability of fusion and routing
- No single fixed fusion, refinement-width, and routing-margin setting dominates across all benchmark families.Datasets vary in the informativeness of names and values, candidate density, and target-schema width.Source: §6 · p. 10
- Learning evidence weights, adapting candidate and refinement widths per query, and routing on the full evidence pattern remain promising directions.The shared configuration provides a stable operating point across datasets.Source: §6 · pp. 10–11
Evaluation and system scope
- The six-family evaluation does not cover every matching regime, and broader testing is needed across domains, model families, and larger target schemas.Some benchmark families have few cases; experiments use one decision backbone and one listwise refinement model, and absolute latency depends on hardware, serving, and batching.Source: §6 · p. 11
7 CONCLUSION
JevNexus combines typed pairwise judgments with schema and instance evidence, reserving listwise reasoning for uncertain cases. Across six benchmark families, it is statistically comparable to Magneto on MRR and Hits@1 while reducing mean online runtime, though Magneto has higher Recall@GT.
7 CONCLUSION
- JevNexus combines typed pairwise judgments with complementary schema and instance evidence, invoking listwise reasoning only for uncertain cases.The design avoids the cost and occasional degradation associated with unconditional generative reranking.Source: §7 · p. 11
- Across six benchmark families, JevNexus is statistically comparable to Magneto on MRR and Hits@1 while reducing mean online runtime by 7.750×, although Magneto attains higher Recall@GT.Evidence fusion provides a consistent benefit, and selective refinement is most useful when evidence sources disagree and the ranking remains uncertain.Source: §7 · p. 11
- Retrieval coverage remains the principal limitation because refinement cannot recover a correct correspondence absent from the candidate set.Future work includes adaptive candidate expansion, cross-query score calibration, and broader evaluation across domains and execution environments.Source: §7 · p. 11