PaperlayerView exact revision on arXiv

Source-linked AI summary

Reflective Compositional Grounding for Natural-Language-to-SPARQL

Isabelle Mohr, John Dujany, Andre Freitas

arXiv:2610.21796v1cs.DB

TL;DR

Natural-language questions can differ from knowledge-graph schemas in wording and conceptual structure, while limited model context constrains schema access. The paper combines compositional grounding, relational query construction, and reflective refinement; across five benchmark/knowledge-graph targets, mean schema recall@10 rises from 0.712 to 0.837.

  • Problem

    Questions can express concepts and relations differently from the target schema, while large schemas may exceed the model’s available context.

    Source: §1 · p. 1

  • Method

    The framework jointly encodes compound expressions, ranks schema candidates using semantic and structural signals, and reflectively revises diagnosed errors in query composition.

    Source: §1 · pp. 1–2

  • Results

    Across five benchmark/knowledge-graph targets, mean schema recall@10 increases from 0.712 to 0.837, and reflective prompt rewriting adds 2.27–3.33 answer-F1 points with retrieval fixed.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The formal analysis specifies conditions for conservative updates that preserve previously validated interpretations across covered runs.

    Source: §2.6 · p. 5

    The experiments do not evaluate persistent mapping memory or cross-query retention, and monotone requirement accumulation alone does not establish semantic compatibility of a changed process.

    Source: §2.6 · p. 5

Abstract

from arXiv · show

Natural-language query generation over large knowledge graphs embodies the classical vocabulary problem in human-systems communication. Questions express concepts through wording and compositions that can differ substantially from the target schema, while limited model context requires selective access to schema information. We present an NL2SPARQL framework that places conceptual compositionality at the center of query interpretation and connects it to the relational composition required for binding and query construction. A fixed concept encoder represents compound expressions jointly, and schema candidates are ranked using semantic similarity, type compatibility, and hierarchy proximity. Reflective refinement critically examines how the grounded meanings are realized through argument bindings, joins, and operator scope. Diagnosed semantic discrepancies guide revisions to the responsible stage. We formalise a monotone refinement discipline that accumulates scoped interpretation requirements and establish conditions under which changes to the composed process preserve previously validated interpretations. Across five benchmark/knowledge-graph targets, mean schema recall@10 increases from 0.712 to 0.837. With retrieval held fixed, reflective prompt rewriting adds 2.27-3.33 answer-F1 points. These experiments evaluate the grounding and refinement mechanisms, while the formal analysis specifies the stronger requirements for conservative adaptation across queries.

1 Introduction

Natural-language questions can differ from knowledge-graph schemas in wording and conceptual structure, while large schemas exceed available model context. The framework combines compositional schema grounding with relational query construction and reflective refinement; its reported experiments improve retrieval and answer quality.

1 Introduction

  • Schema differences can involve conceptual structure as well as wording, and large heterogeneous graphs create contextual ambiguities while their schemas may exceed model context.Relevant schema elements may share little lexical overlap with a question, making selective schema retrieval part of interpretation.Source: §1 · p. 1
  • The investment example distinguishes recognizing “stakes in unlisted companies” as PrivateEquity from connecting the organisation and company through ownership and country constraints.It illustrates how a compound concept’s meaning and its relations must both be supported by retrieved schema context.Source: §1 · p. 1
  • A fixed concept encoder represents compound expressions jointly, and schema candidates are ranked by semantic similarity, type compatibility, and hierarchy proximity.This makes the compound expression the unit compared with schema descriptions before candidates pass to binding and construction.Source: §1 · p. 2
  • Reflective refinement examines argument roles, modifier attachments, joins, and operator scope, then adds a scoped requirement to the diagnosed stage while preserving earlier obligations.The affected stage is recomputed under renewed critique, connecting semantic critique to revisions in query construction.Source: §1 · p. 2, Figure 1 · p. 2
  • The framework formalises compositional representations, contextual interpretation mappings, and staged query processing, with conditions for identifying mappings and a conditional retention theorem.The theorem concerns updates that preserve protected mapping meanings and satisfy compatibility contracts across stage interfaces.Source: §1 · p. 2
  • Across five benchmark/knowledge-graph targets, mean schema recall@10 rises from 0.712 to 0.837, while reflective rewriting adds 2.27–3.33 answer-F1 points with retrieval fixed.These experiments evaluate grounding and refinement mechanisms; the formal analysis addresses stronger conditions for conservative adaptation across queries.Source: §1 · p. 3

2 Proposed Approach

The proposed approach treats query interpretation as both conceptual composition for schema grounding and relational composition for bindings, joins, and operators. It stages retrieval, binding, and construction, then uses reflection to revise diagnosed discrepancies; a separate formal extension states conditions for retaining validated meanings across queries.

2.1 Staged Query Process

  • The query process stages schema grounding, binding, and construction, with each stage receiving the question, relevant prior decisions, and the schema context needed for its task.Grounding retrieves candidate interpretations; binding associates them with entities, literals, and constraints; construction combines bindings into an executable query.Source: §2.1 · p. 3

2.2 Compositional Encoding and Nearest-Neighbour Retrieval

  • A fixed concept encoder jointly represents compound expressions, and structural ranking adds directional type compatibility and hierarchy proximity to embedding similarity.Ranking selects from the retrieved neighbour pool; it cannot recover a schema element absent from that pool.Source: §2.2 · p. 3

2.3 Schema Context and Stage Inputs

  • A query-specific schema view preserves candidate identifiers, descriptions, and relevant type and relation declarations so later stages can interpret candidates in context.The complete schema index and diagnostic records remain outside model context, and further pruning must be recorded because it changes the interpretations available downstream.Source: §2.3 · p. 4

2.4 Relational Binding and Construction

  • Binding assigns entities, values, and constraints to retrieved concepts and relations, while construction determines argument positions, shared variables, joins, and operator scope.An optional symbolic constructor searches restricted property assignments and checks their support against the graph; the default constructor uses an LLM.Source: §2.4 · p. 4

2.5 Reflective Refinement

  • Reflection uses execution results and semantic critique to diagnose whether retrieved meanings were realized through intended bindings and structural relations, then assigns a scoped correction to a stage.The affected stage and its successors are recomputed; for fixed stage inputs, admitted strengthened specifications preserve earlier obligations.Source: §2.5 · p. 4

2.6 Conservative Inference-Time Learning

  • Across questions, a protected bank stores interpretation mappings with applicability scopes and provenance, distinguishing validated interpretations from revisable hypotheses.Mapping identification relies on semantic validation, context separability, and diagnostic traces that justify the proposed mapping, stage, and scope without requiring a unique repair.Source: §2.6 · p. 5
  • Under preservation of protected meanings and interface compatibility contracts, the retention theorem guarantees covered runs retain validated meanings.This is distinct from monotone prompt requirements; the experiments do not evaluate persistent mapping memory or cross-query retention.Source: §2.6 · p. 5

3 Empirical Analysis

The empirical analysis tests compositional retrieval, structural ranking, reflective refinement, and consistency across language-model backbones using benchmark and diagnostic evaluations. The reported results show improved schema recall and answer quality, while the included conditioning analysis cautions against causal interpretation.

3.1 Research Hypotheses

  • The hypotheses test whether compositional encoding, type and hierarchy ranking, reflective refinement, and retrieval across backbones improve schema recall or answer F1.RH1–RH4 separately specify these comparisons against the corresponding baseline or refinement-disabled pipeline.Source: §3.1 · p. 5

3.2 Benchmarks

  • The evaluation covers five benchmark/knowledge-graph targets across four established benchmark families, plus a diagnostic benchmark for composition-related phenomena.QALD-9-Plus is evaluated on both DBpedia and Wikidata; QALD-10 includes live and static reference conditions.Source: §3.2 · pp. 5–6

3.3 Baselines

  • Controlled retrieval comparisons use Flat, Encoder, and Full configurations with generation held constant and reflection disabled; reflection is then tested under Full retrieval.Published results from other systems provide context, but those systems were not rerun.Source: §3.3 · p. 6

3.4 Ablations

  • The ablations distinguish encoder transfer, added structural ranking, prompt rewriting, and retrieval comparisons across backbones, with supporting analyses for gold entities and restricted graph-checked construction.The unreconciled A5 budget summary is excluded.Source: §3.4 · p. 6

3.5 Metrics and Grounding Diagnostics

  • Metrics include answer F1, exact match or execution accuracy, and schema-element recall@k; grounding accuracy and answer F1 are distinct diagnostics.The conditioned comparison reduces variation from named upstream errors but does not identify a causal effect of retrieval or show that the diagnostics coincide.Source: §3.5 · p. 6

3.6 Results

  • Replacing the baseline encoder raises mean schema recall@10 from 0.712 to 0.796, an increase of 8.4 percentage points.The table caption states that Full retrieval improves recall and answer quality on all five targets, with refinement off for the RH1–RH2 comparison.Source: §3.6 · p. 6, Table 1 · p. 6

Appendix E.

Appendix E reports answer-quality gains from prompt refinement and retrieval, alongside structural-ranking improvements. The remaining passages cover related work, the conclusion and limitations, and supporting method and reproducibility material.

Appendix E.

  • Mean recall@10 rises from 0.796 to 0.837 with type compatibility and hierarchy ranking; the full retrieval configuration improves unconditioned answer MF1 over Flat by 5.52–6.53 points on every target.The combined structural score adds 4.1 percentage points, but the separate effects of compatibility and hierarchy distance remain unresolved.Source: Appendix E · p. 7, Table 3 · p. 7
  • 2.27–3.33 answer-F1 points: stage-attributed prompt rewriting improves answers across five targets with full retrieval fixed.Table 4 reports the refinement-minus-no-refinement comparison.Source: Appendix E · p. 7, Table 4 · p. 7
  • Retrieval helps all four tested backbones on both targets, while diagnostics locate the largest retrieval gain in the subsective stratum and report pooled accuracy loss from the restricted solver.Reflection gains persist after conditioning on correct entities and query structure, and with gold entities supplied.Source: Appendix E · p. 7

4 Related Work

  • Related work spans compositional grounding, semantic parsing, schema-guided construction, and reflective correction; this paper combines composition-aware retrieval with staged refinement and formal preservation conditions.The evaluation discussion distinguishes fixed-configuration ablations from external comparisons, which provide context rather than controlled estimates of cross-system superiority.Source: §4 · p. 8

5 Conclusion

  • The conclusion presents a prompt-parameterised grounding, binding, and construction process with a fixed composition-aware encoder and reflective feedback, while treating structured requirement evaluation as future work.The formal retention results are conditional, and reported experiments do not establish their premises.Source: §5 · p. 8

Limitations

  • The reported prompt-rewriting runs lack the admission contract needed for formal retention; ranking also depends on type resolution, schema coverage, candidate-pool recall, relation direction, and depth weight.Repeat runs on two small benchmarks do not establish an ablation noise threshold, and single-run bootstrap intervals measure question-sampling rather than run-to-run variation.Source: §4 · p. 8, Limitations · p. 9

Ethical Considerations; Appendix Guide

  • The work uses public general-domain knowledge graphs without collecting personal or sensitive data; appendices document method, evaluation, results, and literature-review materials.The appendices cover staged interfaces and guarantees, benchmark and model details, configurations and diagnostics, supplementary results, and the expanded literature review.Source: Ethical Considerations · p. 9, Appendix Guide · p. 12

A Staged Query Process and Worked Example; A.1 Accumulated States; A.2 Composition and Hierarchical Ranking; A.3 Construction Scope; A.4 Scoped Feedback and Replay

  • The appendix specifies accumulated stage states, composition and ranking, restricted construction, and scoped feedback replay, with interface responsibilities and conditional properties documented.A requirement combines an applicability condition, behavioural obligation, and diagnostic provenance; replay recomputes the attributed stage and its successors.Source: A Staged Query Process and Worked… · p. 12, §A.1 · p. 12, §A.2 · p. 13

A.5 Dimensions of Mapping Difficulty

The section distinguishes the mapping difficulties addressed by the paper from those left outside its claims, then illustrates the distinction and states guarantees for ranking and restricted construction. Its scope is narrower than full multi-element compositional grounding.

A.5 Dimensions of Mapping Difficulty

  • The section names semantic distance, narrow modifier composition, and schema scale as relevant dimensions, while noting that scale is addressed only indirectly and representation multiplicity and broader scope remain outside evaluation.The diagnostic subsective stratum is only approximately operationalized, and does not separately validate general temporal, comparative, or negative scope.Source: §A.5 · pp. 13–14
  • The paper does not claim to solve full compositional dependence across multiple schema elements and anchors; grounding retrieves candidates independently, leaving connections to query synthesis.The restricted solver searches property-to-slot bindings only for two narrow single-hop query shapes, given a fixed class or entity anchor.Source: §A.5 · p. 14

A.6 A Worked Example

  • The worked example combines lexical mismatch between “stakes in unlisted companies” and PrivateEquity with a same-country constraint requiring shared bindings across ownership and location relations.Resolving the class alone does not answer the question; the query must bind the organisation’s and held company’s headquarters countries to the same variable.Source: §A.6 · p. 14

B.1 Hierarchical Candidate Ranking

  • The ranking guarantee applies to the resolved-type branch; reproducing unresolved metadata cases and the numerical depth-weight function requires configuration details not identified in the summaries.Failure to establish subclass entailment may reflect incomplete metadata rather than genuine incompatibility.Source: §B.1 · p. 15
  • For equal nonnegative similarity, passing candidates nearer in the hierarchy rank at least as highly when the depth weight is non-increasing.A failed candidate’s zero score does not itself remove it from the pool or guarantee a recall improvement; ranking depends on the candidate pool and stated assumptions.Source: §B.1 · p. 15

B.2 Restricted Graph-Supported Construction

  • In the restricted supported search, every returned assignment has a complete branch whose cumulative prefixes passed the endpoint predicate.The predicate provides graph support, not equivalence to the natural-language question.Source: §B.2 · p. 15
  • With a fixed endpoint snapshot, terminating exact checks, exhaustive search without a resource cutoff, and deterministic serialization, the finite search finds any candidate-relative assignment that passes every prefix.The guarantee is limited to supplied candidates, a fixed anchor and supported shape; a semantically wrong property may still pass graph checks.Source: §B.2 · pp. 15–16

B.1 Hierarchical Candidate Ranking

  • The ranking test is directional: it checks a property’s domain against the proposed subject type or admits a class candidate when it is the resolved type or an ancestor.The illustrated type gate can set a failed candidate’s score to zero, while hierarchy distance reduces an ancestor’s score; these are toy values, not measured retrieval results.Source: §B.1 · p. 15, Figure 3 · p. 16

B.3 Scoped Requirements and Targeted Refinement

This section formalizes scoped requirements and targeted replay, then states conditions for conservative process revision. The guarantees concern declared specifications and protected interpretations, not automatic model compliance or semantic correctness.

B.3 Scoped Requirements and Targeted Refinement

  • The formal framework distinguishes requirement accumulation from prompt rendering and states that existing experiments do not empirically certify the structured admission procedure.That procedure has a separate implementation and evaluation obligation.Source: §B.3 · pp. 16–17
  • A scoped requirement is a typed predicate on a stage extension, guarded by when it applies and recorded with diagnostic provenance; a question-specific diagnosis does not license a rule for every question.The encoder remains fixed during refinement, and the named stage interfaces retain their order.Source: §B.3 · p. 17, §C.1 · p. 19
  • Admission requires well-typed predicates, explicit scope and provenance, preservation and consistency checks, and a renderer that preserves the specification; arbitrary natural-language predicates are not assumed decidable.A check restricted to the current input establishes only that local result.Source: §B.3 · p. 18
  • Admitted updates accumulate requirements, so the new feasible extensions satisfy earlier clauses, provided the preservation and consistency checks and faithful-renderer contract hold.This is specification monotonicity, not a guarantee that an LLM obeys a rendered prompt or improves answer accuracy.Source: §B.3 · pp. 17–18
  • After an attributed update, replay retains the prefix before that stage and recomputes it and every successor, while diagnostics are rerun on the latest query.Accepted means the declared checks pass; budget exhaustion, conflict, and no new requirement are distinct stopping statuses, and none certifies natural-language correctness.Source: §B.3 · p. 18

C.1 Contextual Mappings and Protected History

  • Contextual mappings are partial expression-and-context interpretations stored with guards and provenance; validated entries have declared scope, while other entries remain revisable hypotheses.Protected records include the question, context, validated interpretation, query, and trace; validation is not equivalent to obtaining nonempty answers.Source: §C.1 · p. 19

C.2 Identification and Edit Localization

  • Semantic identification is conditional: sound evidence passes the intended mapping, realizability keeps it in the candidate class, and separating evidence excludes every inequivalent candidate.Trace-based localization identifies a proposed edit site but does not certify a unique or correct repair.Source: §C.2 · p. 20

C.3 Local Contracts for a Process Revision

  • Conservative revision requires local interface relations and compatibility contracts across grounding, binding, and construction, with terminal compatibility preserving the validated interpretation.An unchanged prompt alone does not discharge a stage’s contract; stochastic stages require coverage of all covered outputs.Source: §C.3 · pp. 20–21

C.4 Proof and Extent of Retention

The retention result is conditional on local contracts covering every protected record and covered execution. The paper also characterizes its experiment records as retrospective rather than a preregistered or systematic benchmark-selection process.

C.4 Proof and Extent of Retention

  • Under the stated local contracts, every covered updated execution preserves each protected record’s validated composite interpretation and protected mapping meanings.The proof proceeds by stage induction from initialization through grounding, binding, and construction.Source: §C.4 · p. 21
  • Successive accepted updates preserve accumulated records only when each new record is semantically validated and the current procedure’s local contracts are certified for it.Validated samples lacking that coverage remain run-scoped observations; hypotheses outside the protected scope can still change.Source: §C.4 · p. 21
  • The result does not establish universal-domain preservation, guaranteed discovery, or convergence, and the experiments did not evaluate persistent memory, identification certificates, or cross-query retention.For stochastic procedures, every allowed output must be covered; one tested run is insufficient.Source: §C.4 · p. 21

D Benchmark and Baseline Selection

  • The experiment account is retrospective, not a preregistered sampling plan or systematic screening of every available benchmark and system.Benchmark-family selection, question inclusion in a metric, and result retention are separate decisions.Source: D Benchmark and Baseline Selection · p. 21

D.1 Benchmark Families and Evaluation Targets

The evaluation uses selected English NL2SPARQL targets with identified RDF graphs and released splits, supplemented by a separately analyzed composition diagnostic set. The account qualifies the targets’ coverage and representativeness.

D.1 Benchmark Families and Evaluation Targets

  • The public benchmark families were selected for English questions paired with SPARQL or reference answers, an identified RDF graph, and a released split permitting the evaluation target to be specified.The four public families yield five main benchmark/KG targets because QALD-9-Plus is evaluated on two graphs; the diagnostic set is analyzed separately.Source: §D.1 · pp. 21–22
  • The selected targets span QALD grounding on DBpedia and Wikidata, larger-scale LC-QuAD 2.0, and heterogeneous per-domain Spider4SPARQL schemas.The five main-target comparison uses the stated static QALD-10 condition; live and static references are separate conditions.Source: §D.1 · pp. 21, 23
  • The target set is not a random sample of schemas, languages, or deployment settings, and non-English or non-SPARQL/RDF tasks are outside its scope.Other eligible NL2SPARQL benchmarks were not claimed to have been screened and rejected.Source: §D.1 · p. 22
  • Recorded QALD-9-Plus DBpedia and Spider4SPARQL counts differ from the cited upstream releases, and missing release identifiers prevent reconciliation of the subset differences.The account preserves observed counts without inferring filtering rules or unsuccessful queries.Source: §D.1 · p. 22, Table 5 · p. 23
  • The reported record has incomplete or metric-dependent coverage for QALD-9-Plus and Spider4SPARQL, so it cannot establish identical effective denominators or difficulty-independent missingness.Two locally excluded Spider4SPARQL graphs are attributed to data-quality problems, but item IDs and graph-specific fault descriptions are not provided.Source: §D.1 · p. 22
  • The 181-question diagnostic set is a targeted probe of populated graph patterns and composition distinctions, not an estimate of naturally occurring question frequencies.Construction filters precede system comparison; the original subsective-labeled cases remain included with a construct-validity limitation.Source: §D.1 · p. 22

D.2 Baseline and Result Retention

The section defines controlled within-system comparisons that isolate named pipeline components, while external system results are retained only as contextual, non-controlled comparisons. It also excludes result records whose configurations cannot be reconciled.

D.2 Baseline and Result Retention

  • Flat, Encoder, and Full retrieval distinguish the baseline encoder, composition-aware encoder, and added type-and-hierarchy ranking under fixed backbone and prompting settings.A1/A2 test retrieval changes, A4 tests prompt rewriting against refinement disabled, and A6 compares Flat and Full within each backbone.Source: §D.2 · p. 23
  • The auxiliary retry/probe comparison is separate from Full retrieval, and restricted construction and gold-entity variants are treated as distinct diagnostics.The model-selection criteria for A6 are reported separately.Source: §D.2 · p. 23
  • External systems are contextual comparators, not rerun controlled baselines; differing metrics, snapshots, prompts, backbones, and evaluated items preclude a controlled ranking.Values retain their source metrics, and unavailable cells do not imply that a method was tested and failed.Source: §D.2 · pp. 23–24
  • A5 is excluded because its budget assignments conflict with the stated no-refinement configurations, so those records cannot establish an effect of increasing refinement budget.The inconsistent pooled diagnostic retrieval row is also not reconstructed; retained results are not removed merely because their effect is negative.Source: §D.2 · p. 24

D.3 Diagnostic Benchmark Construction

The diagnostic benchmark is built from verified KG patterns and paraphrased questions spanning five composition types, with an acknowledged construct-validity gap in its subsective proxy. The accompanying model sweep and run records have defined coverage limits and do not support broad extrapolation.

D.3 Diagnostic Benchmark Construction

  • The 181-question diagnostic set covers five composition types using live-verified KG patterns and genuine paraphrases; subsective cases use role-discontinuation proxies rather than direct head-dependent modification.The strata are 49 intersective, 29 subsective, 49 relational, 27 modal, and 27 privative questions; the subsective proxy approximates rather than exactly instantiates the semantics.Source: §D.3 · pp. 24–25

E.1 Inclusion and Exclusion Criteria

  • The four-backbone comparison covers paired Flat–Full runs on QALD-10 static-reference and QALD-9-Plus DBpedia targets, testing sensitivity without representing language models generally.Models lacking paired A6 results are outside the comparison, which does not support extrapolation to other benchmarks, conditioned diagnostics, or the preservation theorem.Source: §E.1 · p. 25

E.2 Documented Characteristics

  • Table 6 reports documented backbone capabilities, not recovered run settings; parameter labels do not order computational or reasoning budgets across these systems.The gpt-oss 120B label includes approximately 5.1B active parameters per token, and the models also differ in training, architecture, tokenisation, and decoding.Source: Table 6 · p. 26, §E.2 · p. 26

E.3 Run Configuration and Interpretation

  • The reported conclusion is positive Flat–Full answer-F1 differences within each tested backbone on the two targets, but run details such as snapshots and decoding controls remain unrecorded.The comparison establishes neither a scaling law, equal inference budgets, nor cost efficiency.Source: §E.3 · p. 26

F Evaluation Protocol and Diagnostics

The evaluation uses matched component comparisons and reports both question-sampling uncertainty and repeated baseline-run ranges. Conditioned answer comparisons are restricted to questions passing checks under both configurations and cannot reconstruct direct joint-grounding accuracy.

F.1 Matched Configurations

  • Matched comparisons vary the encoder, structural ranking, reflective rewriting, or backbone while holding the relevant remaining configuration fixed.Structural ranking uses identical initial pools and candidate budgets; a grounding-prompt revision can nevertheless change retrieved candidates.Source: §F.1 · p. 27

F.2 Uncertainty and Repeated Runs

  • The A1–A4 intervals use 10,000 paired question-level bootstrap resamples, while three baseline repeats report DBpedia ranges of 68.00–69.15 MF1 and 56.99–58.06 EA and Wikidata ranges of 66.24–68.46 MF1 and 57.35–59.26 EA.The intervals do not estimate run-to-run variance; the repeat ranges describe only those runs and are not confidence intervals.Source: §F.2 · p. 27

F.3 Conditioned Grounding and Answer Diagnostics

  • Conditioned answer F1 compares both configurations on the shared subset whose entity linking and query structure match references, and selection into that subset limits generalization.The subset’s size and fraction of the full split are reported to expose the conditioning.Source: §F.3 · pp. 27–28
  • Table 8’s executed-answer observations cannot reconstruct joint-grounding accuracy, which requires per-question schema-assignment judgments; answer agreement can occur despite different schema realizations.The backbone results are unconditioned and do not establish invariance of conditioned propagation effects across backbones.Source: §F.3 · p. 28

G.1 Supporting Diagnostics

Supporting diagnostics locate retrieval and answer gains across composition strata, test refinement budgets and retry/probe settings, and derive a bound on solver losses from aggregate results. Several findings are setting-specific, and the comparisons do not establish universal effects or causal mechanisms.

G.1 Supporting Diagnostics

  • Subsective MF1 rises from 62.78 to 71.03 to 83.47, with hierarchy contributing 12.44 of the 20.69-point gain; other strata improve less.The subsective stratum’s proxy operationalization limits the semantic interpretation of this result.Source: §G.1 · p. 28, Table 7 · p. 29
  • On the shared entity-and-structure-correct subset, Full retrieval improves answer F1 by 5.38–6.89 points on every target, while the subsets cover 30.5–65.7% of recorded targets.These are within-subset comparisons and do not automatically extend to excluded questions or identify a causal effect.Source: §G.1 · p. 28, Table 8 · p. 29
  • Recall@10 in the 29-question subsective stratum rises from 0.310 to 0.552 with the encoder and to 0.793 with hierarchy ranking, a combined +48.3 pp, the largest stratum gain.The subsective construction uses role-discontinuation KG properties, an acknowledged proxy limitation.Source: §G.1 · p. 28, Table 7 · p. 29
  • On LC-QuAD 2.0, reflection adds 2.74 points with gold entities, while gold entities add 5.88 points after reflection; the combined gain is 0.59 points below the standalone-gain sum.The interaction difference is descriptive because no uncertainty estimate is reported.Source: §G.1 · p. 28, Table 9 · p. 29

G.2 Retrieval Gains Across Targets

  • Figure 4 compares recall@10 configurations within each target and indicates that the encoder and hierarchy improve recall on every target.Its connecting segments do not interpolate recall across retrieval budgets.Source: Figure 4 · p. 30

G.4 Retry and Grounding Probes

  • Adding grounding probes to retry raises DBpedia MF1 by 8.02 points but lowers diagnostic-benchmark MF1 by 6.50 points; aggregate effects vary by setting and component combination.The comparisons do not identify the query-level cause of the interaction or a universally optimal configuration.Source: §G.4 · pp. 30–31, Table 11 · p. 31

G.3 Refinement-Budget Sweep

  • A refinement budget of 3 captures 95–96% of the eventual budget-10 gain on every target; scores plateau by budget 5 and are unchanged at 10.This pattern is consistent with refinement self-terminating when a round is accepted.Source: Figure 5 · p. 30, §G.3 · p. 30

G.5 External Comparisons and Alternative Scoring

  • External comparison values retain their original metrics and are not used to rank systems because sources differ in metrics, backbones, snapshots, and annotations.Controlled evidence for retrieval comes from within-system comparisons.Source: §G.5 · p. 31

G.7 Case-Level Attribution of Solver-Fired Losses

The case-level analysis attributes solver-fired losses to specific property, label-serialization, and shape-applicability failures. The accompanying review distinguishes compositional meaning, query composition, and process composition, and clarifies the conditions and limits of preservation claims.

G.7 Case-Level Attribution of Solver-Fired Losses

  • Of 41 solver-fired intersective cases, 19 regressed, 3 improved, and 19 were unchanged; coarser-property choices drove most regressions.The solver often chose a live-verified property with a strict or near-superset answer set, while free-text more often emitted the tighter gold property; five cases also failed through missing label triples.Source: §G.7 · pp. 31–32
  • The relational stratum fired on 28 of 49 questions: one regressed, three improved, and 24 were unchanged.The improvements corrected default construction’s added, unverified type restrictions; the single regression involved reconstructing an entity through a missing rdfs:label join.Source: §G.7 · p. 32
  • All five fired modal and privative cases regressed sharply after the solver chose inapplicable single-hop properties instead of the required dct:subject category-membership pattern.Only three modal and two privative questions fired, so they do not support a comparable case-level breakdown.Source: §G.7 · p. 33

H.1 Scope and Selection Procedure

  • The narrative review searched relevant sources through 5 October 2026 but does not claim exhaustive coverage or a preregistered systematic-review protocol.It distinguishes included related work from experimental baselines and does not pool scores across incompatible settings.Source: §H.1 · p. 34

H.2 Compositional Meaning and Representation

  • Linguistic, query, and process composition are distinct; better retrieval of a composed phrase alone cannot establish the interface conditions for preservation.The modifier strata do not constitute a CFQ-style controlled split or establish general compositional generalization.Source: §H.2 · p. 34

H.3 Structured Parsing and Schema Grounding

  • Grounding traces must expose chosen schema elements, attachments, and correction scope so diagnostics can identify the interpretation being revised.Prior parsing and interactive systems expose different intermediate choices, and their differing controllers and evaluation settings do not isolate this paper’s encoder or hierarchy score.Source: §H.3 · p. 35

H.4 Reflection, Prompt Optimization, and Retained Experience

  • The preservation theorem gives sufficient conditions for retaining established query meanings, not evidence that unrestricted reflective prompting satisfies them.The review distinguishes repairing one query from preserving previously learned interpretations across process changes.Source: §H.4 · pp. 35–36

H.5 Validation and the Limits of Correction

  • Execution and schema checks establish properties only within their scope, while answer agreement on one graph does not establish semantic preservation.A universal guarantee requires a sound contract or proof covering its scope; a replay set supplies only finite regression evidence.Source: §H.5 · p. 36

H.6 Evaluation and Implications for This Study

Text-to-SPARQL results depend on model information, candidate generation, and evaluation protocol, while benchmark families test different demands. The paper’s experiments evaluate grounding and refinement, but empirical preservation claims would require a separate ordered-update and replay evaluation.

H.6 Evaluation and Implications for This Study

  • Raw-question generation, gold-entity conditions, demonstrations, and fine-tuned systems require explicit protocols; external scores do not establish a controlled ranking of mechanisms.Benchmark families cover complementary demands, and no single family isolates grounding, structural interpretation, language coverage, and long-term adaptation simultaneously.Source: §H.6 · p. 36
  • Empirical preservation evaluation would require ordered queries, validated-interpretation records, versioned updates, and replay of protected queries after each admitted change.New-query accuracy and retention should be reported separately; current aggregate improvements do not establish the theorem’s identifiability and preservation assumptions.Source: §H.6 · p. 36

I Use of AI Writing Assistance

  • The authors report using Claude to assist with experiment implementation, writing, editing, and LaTeX preparation, and state that they reviewed and edited the content.They take full responsibility for the publication’s content.Source: I Use of AI Writing Assistance · p. 37

Related papers

Loading 2610.21796v1…