{"id":"20dec500-e1cf-4b15-bc67-d2593d9758a2","arxiv_id":"2505.13276","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"CHAD-KG is a new, openly published knowledge graph connecting cultural heritage object metadata with digitisation paradata through a reusable RDF mapping pipeline.","lead":"Researchers built an open knowledge graph that records both what museum objects are and how each object was digitised. The graph, its mapping rules, and the supporting code are publicly available, making the workflow reusable by other heritage digitisation projects.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exact input spreadsheets are not released; only empty templates and the RDF output are described, so the reproducibility claim cannot be independently checked and the source-to-graph mapping cannot be audited.","rationale":"The reader's CONDITIONAL verdict is appropriate. The resource is credible: a public Zenodo dump, SPARQL endpoint, SPUK site, ontology under w3id, and GitHub code independently support the existence and design claims. The weakest point is not the choice of standards or the mapping technology but the closed link between the two tabular datasets and the published graph. The paper describes the collection templates and the output graph but does not state that the actual filled source spreadsheets are available; reference [37] is explicitly the templates. Without them, 'reproducible pipeline' can only mean reusable methodology, not reproducible artifact, and the 'single source of truth' claim cannot be checked for semantic fidelity. This is a concrete, testable gap rather than a reason to reject. It does not change the conditional verdict; it sharpens the condition: the authors should publish the source spreadsheets, or at least a validated sample, and describe the validation performed. The reader's weakest assumption (source spreadsheets complete and accurate) is adjacent but not identical; my concern is that the spreadsheets are not even accessible for inspection, so completeness and accuracy cannot be assessed. No formal verification exists, but none is needed for a dataset paper; the public artifacts are the evidence.","tokens_in":15418,"tokens_out":6622,"duration_ms":63504,"concrete_test":"Download the Zenodo deposit [38] and inspect its contents for the original BM and PD CSV exports. If they are absent, run the published Morph-KGC extension and YARRRML mapping files against the published template/sample data, and compare the generated graph's triple and entity counts and a set of known entity IRIs with the published TTL. If the counts and IRIs cannot be reproduced, the paper should either publish the exact source spreadsheets or explicitly limit the reproducibility claim to the methodology rather than the graph instance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that CHAD-KG demonstrates a reproducible, standards-based pipeline for publishing CHO metadata and paradata. The Morph-KGC extension, mapping files, and templates are public, but the exact BM and PD tabular inputs are not, as described: reference [37] provides the spreadsheet templates with an empty model and a minimal sample, while reference [38] is the TTL graph. Without the actual filled spreadsheets exported to CSV, a third party cannot rerun the pipeline to regenerate the stated 52,080 triples and 14,506 entities, nor can they audit whether the CHAD-AP mapping preserved the intended meaning of each field. Section 3.2 reports collaborative collection in Google Sheets and a custom cleaning/normalisation step in the launch script, but it does not report any completeness check, accuracy audit, or semantic validation, and those checks cannot be performed from the published artifacts alone. This does not falsify the existence of the KG or its design; it means the strongest reproducibility and source-of-truth claims rest on an unverifiable transformation. If the actual CSVs are in fact included in the Zenodo deposit, this concern is resolved; the paper currently gives no indication of that.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents CHAD-KG, a knowledge graph that represents bibliographic metadata and digitisation paradata for cultural heritage objects exhibited in the temporary Aldrovandi exhibition. The graph was generated from two Google Sheets tabular datasets (BM and PD) via a reproducible pipeline based on a Morph-KGC extension, and structured according to CHAD-AP, an OWL application profile reusing CIDOC-CRM, LRMoo, CRMdig, and Getty AAT. The authors report 52,080 triples and 14,506 entities, describe the data model, the materialisation pipeline, and the publishing infrastructure (Zenodo dump, SPARQL endpoint, SPUK static site), and claim that the graph now serves as the main metadata source and single source of truth for the Aldrovandi Digital Twin. The paper's central contribution is a standards-based, openly documented pipeline for publishing heritage object metadata and digitisation paradata as linked data.","tokens_in":15558,"tokens_out":3291,"duration_ms":33452,"significance":"If the claims hold, this is a useful and timely contribution to semantic interoperability in cultural heritage. The paper demonstrates how existing standards (CIDOC-CRM, LRMoo, CRMdig, AAT) can be combined in an application profile and materialised with a declarative mapping tool, and it makes the artefacts available: the ontology, mapping files, code, templates, RDF dump, and SPARQL endpoint are all referenced with DOIs or URLs. There are no equations fitted to the target claim, and the design reuse of CHAD-AP is an acknowledged dependency rather than circularity. The main weaknesses are that the exact input spreadsheets are not released, so the source-to-graph transformation cannot be independently audited, and the paper contains no data-quality validation—only descriptive counts. These issues directly affect the reproducibility and trustworthiness claims that are central to the paper.","major_comments":[{"comment":"The exact BM and PD spreadsheets used to generate CHAD-KG are not released: reference [37] provides only empty templates with a minimal sample, and reference [38] is the final RDF dump. Consequently, a third party cannot rerun the pipeline to regenerate the stated 52,080 triples and 14,506 entities, nor can they audit whether the CHAD-AP mapping preserved the intended meaning of each source field. This is load-bearing for the paper's reproducibility claim. Please deposit the actual exported CSV files used as input (or explicitly state why they cannot be released), and if they are already included in the Zenodo deposit, point the reader to the exact files and paths.","section":"Section 3.2 and Resource availability statement; references [37], [38]"},{"comment":"The paper reports no completeness check, accuracy audit, or semantic validation of the graph against the source spreadsheets or against expert knowledge. Section 4 evaluates the graph only through descriptive statistics—total triples, entity counts, and model-reuse frequencies—which are consistent with the pipeline but do not establish that CHAD-KG is a faithful representation of the metadata and paradata, despite the 'single source of truth' claim made in Sections 3 and 5. Please add a validation step, for example a sampled comparison of entities and properties in the RDF back to the original spreadsheets, a check of mapping edge cases, and a report of discrepancies found and resolved.","section":"Section 4 (Discussion) and Section 3.2"},{"comment":"The cleaning, normalisation, and reshaping performed by the launch script and the user-defined functions are described only at a high level (e.g., 'normalize_and_convert_to_iri', 'split_year_range_to_dates', 'convert_to_aat'). These steps can materially alter the semantic content of the source data—particularly the AAT term mappings and the extraction of identifiers from free-text strings—but the paper does not document the underlying assumptions or transformation rules. Please specify the rules, either in the paper or in a companion documentation file, so that a reviewer or adopter can judge whether the conversions are semantically faithful.","section":"Section 3.2 (User-defined functions and launch script)"}],"minor_comments":[{"comment":"The text contains the grammatical error 'a OWL application profile' and later 'a OWL-encoded application profile'; both should be 'an OWL'.","section":"Section 3.1 and Section 5"},{"comment":"Reference [36] contains rendering artifacts ('Mero&#241, o Pe&#241, Uela'); the bibliographic encoding should be repaired so the author name appears correctly.","section":"Reference list"},{"comment":"The paper should state the exact version and date of the Zenodo dump used for the reported statistics, since the live SPARQL endpoint may evolve; currently the reader cannot tell whether '52,080 triples' refers to the deposited dump, the live endpoint, or both.","section":"Section 3.3 and Section 4"},{"comment":"It would be helpful to clarify whether the controlled vocabularies in the Google Sheets templates were already aligned with AAT during data collection, or whether the AAT mapping is performed only during materialisation by the custom functions; this affects how the AAT reuse counts in Figure 5 should be interpreted.","section":"Section 3.2 and Figure 5"}],"recommendation":"major_revision","confidential_remarks":"This is a resource-and-pipeline paper whose central contribution is the reproducible publication of cultural heritage metadata and paradata. The main gap is verifiability: without the actual input spreadsheets and a validation study, the reproducibility and 'source of truth' claims rest on an unverifiable transformation. The fixes I request—releasing the exact inputs and adding a validation section—are within the scope of a revision and would substantially strengthen the paper. If the journal is not receptive to resource papers with limited analytical evaluation, the editor may also want to consider fit; however, I find the contribution potentially valuable for the semantic-web and digital-heritage communities."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe bottom line: CHAD-KG is a real resource—an openly published knowledge graph of a museum exhibition's objects and their digitisation paradata, with a reproducible pipeline and a working SPARQL endpoint. If you work on cultural heritage linked data, this is worth a look. The paper itself is a solid, slightly over-claimed dataset description.\n\nWhat's actually new: the graph itself (52k triples, ~14.5k entities), the custom Morph-KGC extension with its user-defined functions, and the end-to-end materialisation workflow from two Google Sheets to RDF. The CHAD-AP ontology appeared in their ISWC paper last year, so the model is not new; the contribution is the concrete instantiation and the tooling. The authors are transparent about reusing existing standards (CIDOC-CRM, LRMoo, CRMdig, AAT) and they explain their design choices well. The publishing setup—Zenodo, SPARQL, HTML static site via SPUK—is genuinely FAIR and praiseworthy.\n\nThe soft spots are in the evaluation, or rather the lack of it. There is no accuracy or completeness check that the graph faithfully represents the source spreadsheets, no semantic validation, no user study. Given that they call the KG the 'source of truth' for the digital twin, that's a meaningful gap. The stronger issue is reproducibility: the exact filled spreadsheets are not released; only empty templates and the RDF output. That means nobody can actually rerun the pipeline to regenerate the stated triple count or audit the field-level mapping. The paper says the templates are available for replicating the collection process, but that's not the same as providing the actual input data. If the CSVs are in the Zenodo deposit, the paper should say so explicitly. This is fixable, and it doesn't undermine the existence of the graph, but it does undercut the 'reproducible pipeline' headline claim.\n\nThe impact statements in Section 4 ('fills a critical gap') are mostly aspiration. There is no evidence of reuse outside the project, no external users, no quantitative evaluation. That's typical for a resource paper, but the phrasing could be tempered.\n\nWho should read this: digital humanists and semantic web practitioners building similar digitisation workflows. They'll find a good example of how to combine established ontologies with a pragmatic mapping tool. It deserves a proper peer review at a venue like Semantic Web or a DH journal, with the validation and data-availability issues addressed.\n\nMy verdict: accept with revisions, not desk reject.\n\n- R","headline":"A genuinely open cultural heritage KG with a solid pipeline, but the missing source data and lack of validation mean the reproducibility and impact claims run ahead of the evidence.","tokens_in":16153,"tokens_out":2443,"would_cite":false,"duration_ms":23598,"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":"This paper shows that object metadata and digitisation paradata can be combined in one standards-based knowledge graph, and demonstrates it with CHAD-KG, a 52,080-triple graph now serving as the single metadata source for the Aldrovandi…","keywords":["RDF","Knowledge Graph","Cultural Heritage","Linked Open Data","Digitisation","Data Modelling","Data Materialisation","Paradata"],"falsifier":"Run the public SPARQL endpoint against the recorded digitisation process of a specific exhibited object—say, tracing its acquisition step, software activities, actors, tools, and dates—and compare the returned triples with the digitisation team's own project records and with the original spreadsheet rows; any missing, altered, or misdated step would show that the graph does not faithfully preserve the intended metadata and paradata.","tokens_in":15151,"feed_emoji":"🏛️","tokens_out":5747,"duration_ms":50544,"temperature":0.7,"pith_summary":"This paper is trying to establish that cultural heritage object metadata and digitisation paradata—the record of how a digital surrogate was acquired and processed—can be published together as a single interoperable knowledge graph using existing semantic-web standards. It presents CHAD-KG, a graph of 52,080 RDF triples and 14,506 entities describing the objects of a temporary exhibition on Ulisse Aldrovandi and the people, activities, tools, and time-spans behind their digitisation. The graph is generated by a reproducible pipeline from two tabular spreadsheets, is released openly, and now functions as the main metadata source for the exhibition's Digital Twin. The value of the claim, if true, is that museums can capture not just what an object is but how its digital replica came to be, in a machine-readable form that other institutions can reuse.","feed_headline":"52,080 triples link museum objects to their digitisation history","feed_subtitle":"A standards-based graph now serves as the single metadata source for the Aldrovandi Digital Twin.","key_machinery":"The carrying object is CHAD-AP, an OWL application profile built on CIDOC-CRM and its extensions. It has two modules: the Object Module, using CIDOC-CRM and LRMoo, models a cultural heritage object through a Work-Expression-Manifestation-Item layering, with titles, creation activities, subjects, type, license, and curation; the Process Module, using CRMdig, models digitisation as a sequence of activities—an acquisition step and later software-execution steps—each with input/output data objects, actors, tools, techniques, and time-spans. The materialisation pipeline turns two spreadsheet templates into RDF according to this profile through reusable mapping rules and custom functions that normalise identifiers, split multi-valued fields, convert dates, and map techniques to Getty AAT codes.","core_discovery":"The central discovery is the working integration of three things: an OWL application profile (CHAD-AP) that reuses CIDOC-CRM, LRMoo, CRMdig, and Getty AAT to describe both objects and digitisation processes; a materialisation pipeline that converts two collaboratively edited spreadsheets into RDF through declarative mapping rules plus custom functions; and the resulting graph CHAD-KG, which is published under CC0 and queried as the source of truth for the Aldrovandi Digital Twin. The paper argues that this combination makes digitisation paradata first-class data rather than informal notes.","pith_inferences":["The authors state the graph is the main metadata source; a corollary they do not spell out is that the graph inherits every error in the source spreadsheets, so a validation step becomes critical as the graph grows.","Because the templates are in Italian and the workflow weaves human expert entry with automated mapping, a natural next test is to map the same templates for a different exhibition or institution to see where the fixed mapping rules break.","The missing aspects the authors list—human-readable labels, physical dimensions, materials—suggest that CHAD-AP will need to grow before it can cover museum records beyond this exhibition.","One can imagine extending the Process Module to record conservation actions and loans alongside digitisation, turning the graph into a full life-cycle record of each object."],"forward_implications":["If the central claim holds, the same CHAD-AP profile and mapping templates can be applied to new collections without rewriting the pipeline.","The graph gives the Digital Twin a single, queryable source of truth, so any correction to object or process data propagates to all views of the twin.","The SPARQL endpoint and CC0 dump make the exhibition's metadata and digitisation history openly reusable by other projects.","Other digitisation case studies in the same programme can integrate their data into the same graph, enabling cross-collection comparison of digitisation workflows."],"supporting_citations":[{"why":"supplies CHAD-AP, the application profile that defines the model used by CHAD-KG.","marker":"[8]"},{"why":"is the published CHAD-KG dataset itself, the central artifact described in the paper.","marker":"[38]"},{"why":"provides the RML-based materialisation engine that the paper's pipeline extends.","marker":"[3]"},{"why":"defines RML, the mapping language used to convert tabular data into RDF.","marker":"[13]"},{"why":"describes the Aldrovandi Digital Twin for which CHAD-KG serves as the metadata source.","marker":"[4]"},{"why":"defines CIDOC-CRM, the foundation of the Object Module.","marker":"[14]"},{"why":"defines CRMdig, the basis of the Process Module for digitisation activities.","marker":"[17]"},{"why":"supplies Getty AAT controlled vocabularies used for types, roles, and techniques.","marker":"[24]"}],"fun_headline_variants":["CHAD-KG links museum objects to their digitisation provenance","52,080 triples make digitisation history first-class data","Open knowledge graph turns spreadsheet digitisation logs into RDF","Standards-based graph powers Aldrovandi digital twin","Digitisation paradata get a queryable home in CHAD-KG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two source spreadsheets are complete and accurately reflect what the digitisation team did, and that the mapping rules preserve that meaning; no completeness, accuracy, or semantic validation of the source data is reported.","fun_headline_variants_meta":{"raw":{"variants":["CHAD-KG links museum objects to their digitisation provenance","52,080 triples make digitisation history first-class data","Open knowledge graph turns spreadsheet digitisation logs into RDF","Standards-based graph powers Aldrovandi digital twin","Digitisation paradata get a queryable home in CHAD-KG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1531,"prompt_tokens":903,"completion_tokens":628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":543}},"tokens_in":519,"tokens_out":628,"duration_ms":6364,"temperature":1.0,"reasoning_tokens":543,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:16:06.437242+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the public SPARQL endpoint against the recorded digitisation process of a specific exhibited object—say, tracing its acquisition step, software activities, actors, tools, and dates—and compare the returned triples with the digitisation team's own project records and with the original spreadsheet rows; any missing, altered, or misdated step would show that the graph does not faithfully preserve the intended metadata and paradata.","supporting_citations":[{"cited_title":"In: Demartini, G., Hose, K., Acosta, M., Palmonari, M., Cheng, G., Skaf-Molli, H., Ferranti, N., Hernández, D., Hogan, A","cited_arxiv_id":null,"evidence_quote":"supplies CHAD-AP, the application profile that defines the model used by CHAD-KG."},{"cited_title":"Ldow 1184 (2014)","cited_arxiv_id":null,"evidence_quote":"defines RML, the mapping language used to convert tabular data into RDF."},{"cited_title":"AI Magazine 24(3), 75 (Sep 2003)","cited_arxiv_id":null,"evidence_quote":"defines CIDOC-CRM, the foundation of the Object Module."},{"cited_title":"In: 3rd USENIX Workshop on the Theory and Practice of Provenance(TaPP11)(2011), https://www.usenix.org/legacy/events/tapp11/ tech/final_files/Doerr.pdf","cited_arxiv_id":null,"evidence_quote":"defines CRMdig, the basis of the Process Module for digitisation activities."},{"cited_title":"Art Documentation: Journal of the Art Libraries Society of North Amer- ica 29(1), 67–72 (Apr 2010)","cited_arxiv_id":null,"evidence_quote":"supplies Getty AAT controlled vocabularies used for types, roles, and techniques."}],"review_version":1}