{"id":"bf6bfbc0-6732-440f-b538-4aa2ef0e195f","arxiv_id":"1908.03171","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A state-of-the-art review that formalizes ontology completion and debugging as an abductive repair problem and maps existing methods to that framework.","lead":"This paper surveys how ontologies, the structured vocabularies behind many data systems, are repaired when they contain wrong or missing information. It wraps the repair step in a formal abduction framework and identifies open problems for both single ontologies and networks of ontologies.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Definition 1 never couples the oracle to M and W; repair guarantees rely on an informal all-knowing-oracle assumption that is not part of the formalization.","rationale":"The reader identifies the oracle assumption as the weakest point, and I agree that oracle reliability is the central vulnerability. My pass sharpens this: the specific defect is not only that domain experts make mistakes, but that Definition 1 as written never connects the oracle to the given defect sets M and W. The all-knowing-oracle case in Section 4.1.2 is an informal narrative that supplies the missing link, so the formalization is less self-contained than the survey's framing suggests. I do not think this changes the verdict, because the paper is a survey rather than a soundness-critical technical system, it explicitly discloses the oracle limitation in multiple places (Section 4.1.2, Section 7.1, Section 7.2), and the descriptive claims about the state of the art do not depend on resolving this gap. The detailed examples and the explicit enumeration of open problems are independently valuable. The recommended action is therefore unchanged: accept the survey while noting that Definition 1 should be tightened, for example by adding the oracle-validity conditions on M and W, if the formalization is to be used as a specification for future algorithms.","tokens_in":34486,"tokens_out":7986,"duration_ms":90668,"concrete_test":"Augment Definition 1 by adding the two missing oracle-validity conditions, namely ∀ψm∈M:Or(ψm)=true and ∀ψw∈W:Or(ψw)=false, and then re-verify the example repairs R1-R8 in the appendix and the state-of-the-art classification in Sections 5-6. If all cited systems that the paper counts as fitting the framework still satisfy the strengthened definition, the informal all-knowing-oracle assumption is benign and the formalization is easily repaired; if some previously included approaches fail the strengthened conditions, the survey's uniform-framework claim is narrower than stated and the prose in Section 4.1.2 is doing work that Definition 1 does not.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The formalization's central repair definition has a gap: Definition 1 constrains the oracle only on the added set A and the deleted set D (conditions (i)-(ii)), while conditions (iv)-(v) simply require that the given missing axioms M become derivable and the given wrong axioms W become non-derivable. Nothing in the definition requires Or to have validated M or W. The paper's Section 4.1.2 asserts that with an all-knowing oracle we can check that every axiom in M is oracle-true and every axiom in W is oracle-false, and can then remove falsely identified defects, but this check is external to Definition 1 rather than part of the formal problem statement. Moreover, the all-knowing property itself is described only informally ('the oracle's answer is always correct'), never as an axiom of the framework. As a result, all of the preference relations (Definitions 2-3) compare repairs by entailment of oracle-true and oracle-false statements, so the entire formal apparatus inherits the correctness of an assumption that is not encoded. The central claim that the repairing step is 'formalized' as an abduction problem therefore depends on an unstated correctness condition that does substantial logical work: it is what guarantees that a repair satisfying (iv) cannot derive a false missing axiom and that (v) cannot suppress a true wrong axiom. This is a real soft spot, but it is a precision gap rather than a hidden fatal flaw, because the paper explicitly acknowledges oracle fallibility and the need for expert validation in Sections 4.1.2 and 7.1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a survey of ontology debugging and completion focused on the repair step. It introduces a formalization of a Complete-Debug-Problem CDP(T,C,Or,M,W), where a repair is a pair (A,D) of TBox axiom sets to add and delete such that: added axioms are declared true by an oracle Or, deleted axioms are declared false, the resulting TBox is consistent, all missing axioms M are entailed, and all wrong axioms W are not entailed (Definition 1). It then defines preference relations among repairs—more complete, less incorrect, subset minimality, and combined skyline optimality—and uses this framework to organize the state of the art for single ontologies and ontology networks, including correctness-only, completeness-only, and combined approaches. The paper concludes with open research problems in theory, algorithms, and user support, and an appendix works through a Galen-inspired EL example verifying several repairs against the five conditions.","tokens_in":34751,"tokens_out":10883,"duration_ms":111376,"significance":"The proposed framework is a useful unifying device: it makes explicit that repair involves both addition and deletion, it connects repair to abduction, and it provides preference notions that prior surveys lacked. The worked examples are careful, and the appendix verifies each candidate repair against Definition 1, which is a strength. The survey's descriptive claims are supported by references, and the separation of debugging and completion is convincingly documented. The paper does not claim machine-checked proofs or experimental results, which is appropriate for a survey; its value lies in the formalization and organization of the field.","major_comments":[{"comment":"Definition 1 constrains the oracle only on the added set A and the deleted set D (conditions (i)–(ii)); conditions (iv)–(v) apply to M and W without any requirement that Or has classified these axioms. The 'all-knowing' case is introduced informally in §4.1.2, and the statement that one can 'without loss of generality' assume that M-axioms are really missing and W-axioms are really false is an external meta-assumption rather than a consequence of Definition 1. As written, therefore, a repair can satisfy (iv) by entailing an axiom that is in fact false according to the domain (if M contains a false positive), and can satisfy (v) by making a true entailment non-derivable (if W contains a false negative); the preference relations in Definitions 2–3 inherit this issue because they compare entailment sets using a possibly fallible Or. I recommend adding explicit conditions to the definition of CDP or to Definition 1 (e.g., ∀ψ∈M: Or(ψ)=true and ∀ψ∈W: Or(ψ)=false, possibly as a separate 'validated defects' assumption) and stating formally what 'all-knowing' means (e.g., Or(ψ)=true iff ψ holds in the intended domain).","section":"§4.1.1 (Definition 1) and §4.1.2"}],"minor_comments":[{"comment":"In the sentence 'Therefore, in Def. 1, ∀ψm∈M:Or(ψp)=true, W=∅ and D=∅', the variable ψp should be ψm.","section":"§5.2"},{"comment":"Definition 1 includes C in the problem signature, but C is never used in any of the conditions; either remove it or state explicitly that it is part of the signature for notational uniformity.","section":"§4.1.1 (Definition 1)"},{"comment":"Definition 1 does not explicitly require D⊆T or A∩D=∅ in general; the examples and the identity (T∪A)\\D=(T\\D)∪A in §4.1.2 assume A∩D=∅. Please state the intended domain of A and D.","section":"§4.1.1 (Definition 1) and §4.1.2"},{"comment":"The color coding and the T/F labels are helpful, but the captions do not explain that blue marks missing axioms and red marks wrong axioms; making the captions self-contained would improve readability.","section":"Figures 2 and 3"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the paper is within the scope of the journal, and the formalization gap in Definition 1 is fixable in revision; I do not see grounds for rejection. The author's own systems are cited prominently, but they are directly relevant prior work, so I do not regard the citation pattern as inappropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read it. First, this is a survey with a light formalization, not a new result with algorithms or experiments. Its value is the uniform Complete-Debug-Problem (CDP) framework and the preference relations that let you compare repair approaches on a common footing. Second, the all-knowing oracle assumption is acknowledged in prose but left out of Definition 1, and that is a real precision gap, but the paper earns the informal version; it is not a hidden fatal flaw.\n\nWhat the paper does well: the definitions are clear, the running examples are worked through step by step (including the Galen-inspired appendix, where each repair is checked against all five conditions of Definition 1), and the state of the art is organized sensibly. Separating debugging from completion and showing that almost no system does both with preferred-repair guarantees is a genuinely useful observation. The author cites his own prior work where it is the actual state of the art (abduction-based completion, RepOSE), so the self-citation pattern is not a problem.\n\nThe stress-test note about the oracle is accurate: conditions (i)–(ii) constrain the oracle on A and D, but conditions (iv)–(v) only require entailment of M and non-entailment of W, with no formal link that validates M and W against the oracle. Section 4.1.2 does state that with an all-knowing oracle you can check M and W and remove falsely identified defects, so for that ideal case you can assimilate M and W into the oracle-supervised repair. But the all-knowing property itself is never made an axiom of the framework, and the preference relations in Definitions 2–3 quantify over all axioms while relying on the oracle's verdict for each. For a limited or fallible oracle, those definitions get murkier. It's a moderate formal nit: the paper could tighten the CDP definition by adding an explicit oracle-correctness condition for M and W, or by defining the all-knowing oracle as part of the framework rather than a comment. As-is, the framework is still coherent, but a careful reader will notice that the formalization and the main narrative about repair guarantees are slightly out of sync.\n\nWho this is for: graduate students entering ontology repair, or researchers in semantic technologies who want a unified problem statement before diving into the literature. It is not for someone looking for new computational results or empirical comparisons, because there are none.\n\nRecommendation: this deserves a serious referee. I would accept it for a venue like the Journal of Web Semantics or Applied Ontology, and the revision request is straightforward: make the oracle's correctness role explicit in Definition 1 or the problem statement, and clean up the preference definitions for non-ideal oracles.","headline":"A solid survey whose CDP framework gives the field a common vocabulary; the oracle gap is a precision issue, not a fatal one.","tokens_in":35258,"tokens_out":2056,"would_cite":false,"duration_ms":23696,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One repair definition unifies ontology completion and debugging.","keywords":["ontology repair","ontology debugging","ontology completion","abductive reasoning","description logics","TBox","preference relations","complete-debug problem"],"falsifier":"Take a repair produced by an axiom-weakening tool and try to express it as a pair ($A$,$D$) of whole axioms under any oracle; if there is an ontology where the weakened axiom cannot be split into oracle-true additions and oracle-false deletions that keep all missing axioms derivable and all wrong axioms underivable, then Definition 1 does not cover every repair method it claims to unify.","tokens_in":34272,"feed_emoji":"🧩","tokens_out":7343,"duration_ms":83439,"temperature":0.7,"pith_summary":"This paper attempts to establish that the repair phase of ontology engineering—the phase after defects have been detected—can be stated as one uniform abduction problem rather than a collection of ad-hoc tasks. In the formalization, called the Complete-Debug-Problem, an ontology is a description-logic TBox, the user supplies sets of missing and wrong axioms, and a repair is a pair consisting of axioms to add and axioms to delete such that the added axioms are true according to a domain-expert oracle, the deleted ones are false, and the resulting TBox is consistent and derives exactly the right statements. Situating prior work inside this definition shows that debugging and completion have usually been studied separately, that a naive combination of a debugging step and a completion step does not generally yield a repair, and that most existing methods assume an all-knowing oracle. A sympathetic reader would care because the framework gives a common yardstick for comparing tools, makes explicit which assumptions each method makes, and clarifies what open algorithmic problems must be solved before ontologies can be kept both correct and complete.","feed_headline":"One repair definition unifies ontology completion and debugging","feed_subtitle":"A repair adds only correct axioms, deletes only false ones, and keeps the ontology consistent—but few tools do both.","key_machinery":"The load-bearing object is the Complete-Debug-Problem (CDP) and its repair relation from Definition 1, together with the oracle $Or$ that answers true or false for each TBox axiom and represents the domain expert. The oracle is the only channel through which domain correctness enters; without it, 'missing' and 'wrong' are just syntactic labels. The framework's other machinery is a family of preference relations over repairs—more complete, less incorrect, and subset minimal—and their combinations, which let the survey classify algorithms by which kind of optimal repair they can produce, such as maximally complete, minimally incorrect, or skyline-optimal. The CDP carries the argument by turning every repair task into the same abductive schema, making the comparison of methods a matter of checking which conditions and preferences each method realizes.","core_discovery":"The central claim is Definition 1: for a TBox $T$, atomic concepts $C$, an oracle $Or$ returning true or false for every axiom, a finite set $M$ of missing axioms and a finite set $W$ of wrong axioms, a repair of CDP($T,C,Or,M,W$) is any pair ($A$,$D$) of finite axiom sets such that every axiom in $A$ is oracle-true, every axiom in $D$ is oracle-false, the TBox obtained by deleting $D$ from and adding $A$ to $T$ is consistent, every axiom in $M$ is derivable from the repaired TBox, and no axiom in $W$ is derivable from it. The author then reads the literature through this lens: correctness-only debugging is the special case with $M$ empty and $A$ empty, completeness-only completion is $W$ empty and $D$ empty, and work that does both at once is rare and mostly restricted to lightweight is-a structures. On this reading the field's main gap is not detection but the combined repair problem and the algorithmic guarantee of preferred repairs.","pith_inferences":["If the CDP formalization became the standard interface for repair tools, benchmark evaluations could be built around a shared tuple ($T$,$M$,$W$,oracle answers), letting any proposed repair algorithm be scored against the same oracle judgments; current surveys compare systems only via their own benchmarks.","The oracle dependence suggests a robustness program the paper only touches: repairs could be defined relative to probability-of-correctness estimates or to multiple experts, so that add and delete decisions are made under uncertainty rather than by an all-knowing judge.","Axiom-level add/delete granularity may be the wrong cut for some real repairs, which weaken or rewrite axioms; the paper notes these can be encoded as delete-plus-add, but this encoding may hide a finer-grained repair problem worth formalizing separately."],"forward_implications":["Every debugging-only or completion-only method discussed is a special case of CDP repair, which is why the framework can compare them uniformly.","A repair for the combined problem cannot in general be obtained by doing a completion step and then a debugging step, or vice versa, since additions can re-enable wrong derivations and deletions can remove needed support.","Preference relations distinguish good repairs: maximally complete, minimally incorrect, and subset-minimal are different goals, and existing systems often realize only some of them.","The open problems follow directly: algorithms for combined repair with guarantees, complexity results for completion and preferred repairs, and extensions to oracles that answer 'unknown'.","Ontology networks can be treated as single TBoxes by translating all mappings into axioms, so the same CDP definition applies there as well."],"supporting_citations":[{"why":"Supplies the description-logic semantics, models, TBoxes, and subsumption that Definition 1 is built on.","marker":"[8]"},{"why":"Origin of tableau-based MUPS computation for incoherent terminologies, grounding the glass-box debugging line.","marker":"[10]"},{"why":"Presents a unified approach for debugging and completing taxonomies and their alignments, the main existing case of combined repair.","marker":"[25]"},{"why":"An earlier survey of debugging approaches used as the baseline showing that debugging methods predate completion work.","marker":"[29]"},{"why":"Provides criteria and background-knowledge notions that the paper adopts for comparing repair approaches.","marker":"[30]"},{"why":"Defines MUPS, MIPS, and pinpointing via hitting sets, the core machinery for correctness repair by deletion.","marker":"[32]"},{"why":"Gives the method for computing all justifications, the standard basis for explanation and repair in debugging.","marker":"[33]"},{"why":"Formulates repair of incomplete EL ontologies as abduction with completeness and subset preferences, the direct predecessor of the CDP.","marker":"[56]"}],"fun_headline_variants":["Rare tools both complete and debug ontologies jointly","Unified repair definition: add true axioms, delete false","Few tools jointly add correct and remove wrong axioms","Combined repair of missing and wrong axioms is rare"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a domain expert can serve as an oracle that returns the correct true or false judgment for every TBox axiom, because the definition of a repair lets axioms be added only on 'true' answers and deleted only on 'false' answers.","fun_headline_variants_meta":{"raw":{"variants":["Rare tools both complete and debug ontologies jointly","Unified repair definition: add true axioms, delete false","Few tools jointly add correct and remove wrong axioms","Combined repair of missing and wrong axioms is rare"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000783,"raw_usage":{"total_tokens":3420,"prompt_tokens":869,"completion_tokens":2551,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":2497}},"tokens_in":485,"tokens_out":2551,"duration_ms":17262,"temperature":1.0,"reasoning_tokens":2497,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:20:55.911266+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a repair produced by an axiom-weakening tool and try to express it as a pair ($A$,$D$) of whole axioms under any oracle; if there is an ontology where the weakened axiom cannot be split into oracle-true additions and oracle-false deletions that keep all missing axioms derivable and all wrong axioms underivable, then Definition 1 does not cover every repair method it claims to unify.","supporting_citations":[{"cited_title":"Lambrix, V","cited_arxiv_id":null,"evidence_quote":"Presents a unified approach for debugging and completing taxonomies and their alignments, the main existing case of combined repair."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"An earlier survey of debugging approaches used as the baseline showing that debugging methods predate completion work."},{"cited_title":"Haase, G","cited_arxiv_id":null,"evidence_quote":"Provides criteria and background-knowledge notions that the paper adopts for comparing repair approaches."},{"cited_title":"Kalyanpur, B","cited_arxiv_id":null,"evidence_quote":"Gives the method for computing all justifications, the standard basis for explanation and repair in debugging."},{"cited_title":"Wei-Kleiner, Z","cited_arxiv_id":null,"evidence_quote":"Formulates repair of incomplete EL ontologies as abduction with completeness and subset preferences, the direct predecessor of the CDP."}],"review_version":1}