Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Data2Concept2Text: An Explainable Multilingual Framework for Data Analysis Narration

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that natural language narration of data can be produced by a transparent, rule-based rewriting system: concept trees are rewritten through six stratified stages into multilingual sentences, with every rewrite recorded…

desk verdict A well-structured, genuinely modular explainable NLG pipeline that deserves referee time, but the 'perfect semantic preservation' claim needs formal backing or a major tone-down. read the letter →

arxiv 2502.09218 v1 pith:5AMZKGXB submitted 2025-02-13 cs.LO cs.AI

Flavio Bertini (UNIPR) , Alessandro Dal Palù (UNIPR) , Federica Zaglio (UNIPR) , Francesco Fabiano (NMSU) , Andrea Formisano (UNIUD) This is my paper · ORCID
classification cs.LOcs.AI
keywords explainableAInaturallanguagegenerationtreerewritinglogicprogrammingProlog/CLPmultilingualconcept-to-textdatanarration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that the second half of an explainable data-narration pipeline, turning abstract concepts into natural language, can be built without statistical language models. It claims that a uniform Prolog-based rewriting system, which transforms concept trees through six stratified stages, is general enough to handle complex relations and modular enough to switch languages. A sympathetic reading: if the system works as described, it gives a transparent, verifiable path from data to fluent sentences, with each generated sentence carrying its own rewrite trace. The authors demonstrate this by enumerating tens of thousands of paraphrases of one conference-narration concept in English and Italian, and by comparing against LLMs that skip attributes and cannot guarantee semantic integrity.

What carries the argument

The load-bearing mechanism is the stratified tree-rewriting pipeline driven by Prolog clauses of the form rule(Lang, Type, Name, Tree, RewTree), applied by breadth-first traversal until no rule in that stage fires. Each stage rewrites the concept tree into a new tree; Equivalent Concepts introduces semantic variants, Concept2Structure builds noun, verb, and prepositional phrase skeletons annotated with info/4 phrase tags, Structure2Grammar replaces concepts with lexemes, Coordination unifies gender and number variables, Inflection orders words by a language-specific constraint satisfaction problem, and Syntax fixes local word-level issues like contractions. The explanation mechanism is an auxiliary tree, RuleTree, isomorphic to the rewritten tree and holding the rule Name applied at each node, so the entire derivation from concept to sentence is inspectable.

What would settle it

Take the pipeline's English rule set and the conference-narration concept used in the paper, generate the full sentence list, and have independent native speakers judge whether each sentence is equivalent to the original concept; any generated sentence that drops, adds, or reverses a relation would refute the strict-transitivity claim. Equivalently, probe a single equivalence rule, such as equiv_interval, across all unit-of-measure classes to find a composition where the 'interval of time' replacement does not hold.

Watch

Extended reading notes

Core claim

The central claim is that a stratified, rule-based rewriting system can carry concept trees to well-formed natural language while preserving meaning exactly, and that the rewriting itself is the explanation: each applied rule is recorded in a trace tree isomorphic to the rewritten tree. The paper states that 'a uniform rule-based rewriting system to drive the Concept2Text process is general and modular,' and that equivalent rewritings 'preserving semantics perfectly' arise from 'the strict transitivity of the applied equivalences.' The system takes concepts as trees of classes and relations, applies language-independent stages (Equivalent Concepts, Concept2Structure, Coordination) and language-dependent stages (Structure2Grammar, Inflection, Syntax) to fixpoint, and flattens the final tree into a sentence. The claim includes multilingual generality, demonstrated with English and Italian rule sets producing thousands of distinct but semantically equivalent sentences from a single input concept.

Load-bearing premise

The system assumes that every hand-written rewriting and equivalence rule preserves meaning and produces grammatical structure; if even one rule encodes a semantic shift, the promised perfect preservation fails for every sentence built through it.

Editorial extensions

If this is right

  • Medical and other safety-critical reports could be generated by a pipeline whose every sentence carries an audit trail of which rules produced it.
  • Adding a new language reduces to writing rule sets for three language-dependent stages, leaving semantic-equivalence and coordination logic untouched.
  • Because variants are enumerated explicitly, the system can offer a user a controlled menu of paraphrases with the same meaning, rather than a single opaque output.
  • The recorded rule trace can double as a machine-readable justification for transparency requirements such as those in the EU AI Act.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial: the rule sets themselves are the real specification of 'semantic equivalence,' so they could be treated as testable artifacts; automatically searching for input concepts where two rules compose into a construction that native speakers judge non-equivalent would expose concrete failure cases.
  • Editorial: the same architecture could be hybridized, using a statistical model for lexical choice or surface polish while keeping the rule trace as the semantic scaffold, without giving up the explainability guarantee.
  • Editorial: the strict-transitivity claim implies that if any single rule is meaning-changing, the guarantee fails globally, making compositional, per-rule verification a natural next step the paper does not spell out.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents Concept2Text, a Prolog/CLP-based rewriting system that transforms concept trees, expressed in terms of classes and relations, into natural language sentences. The pipeline consists of six rewriting stages (equivalent concepts, concept-to-structure, structure-to-grammar, coordination, inflection/sorting, and syntax), and the system supports English and Italian. The authors demonstrate the pipeline on a concept describing the ICLP 2024 call for papers and on a data-analysis example (Scopus xAI publications), enumerating thousands of sentence variants. They claim the system is fully explainable, modular, and preserves semantics perfectly through transitivity of the applied equivalence rules.

Significance. If the semantic-preservation claim can be substantiated, the system offers a genuinely transparent, multilingual alternative to LLM-based data narration, relevant for safety-critical and accessibility applications. The multi-stage tree-rewriting architecture is a clean design, and the explicit RuleTree provenance for each rewriting step is a promising mechanism for explainability. The enumeration of 13,000+ English and 3,200+ Italian sentences, with a downloadable sentence list, is a concrete and reproducible demonstration of the pipeline's generative capacity, which is a credible strength. However, the core claim that all generated variants are semantically equivalent to the input is currently supported only by informal inspection of hand-written rules, not by a formal semantics or an external evaluation.

major comments (3)
  1. [§3.1.2 and §4] The central claim that the rewriting process "preserves semantics perfectly" (Section 4, paragraph beginning "It is noteworthy ...") rests on the unverified assumption that every hand-written equivalence rule in the equiv_concept stage is a genuine semantic equivalence. Section 3.1.2 presents rules such as equiv_interval and measure_range without a formal definition of the semantics of the concept-tree language, and the transitivity argument in Section 4 can only be sound if each rule is a single-step equivalence. The paper's own validation, "We locally checked each rewriting rule behaviour" and "We manually reviewed some samples" (Section 4, paragraphs under "General concept narration"), is informal and does not cover the full rule set. To support the claim, the authors should either provide a model-theoretic or denotational semantics for concept trees and prove each rule sound with respect to it, or replace the "perfectly" claim with a weaker claim supported by a systematic external evaluation of semantic equivalence.
  2. [§3, "Concepts" and §5] The manuscript states "From now on, we assume to work with a simple spanning tree of the original (sub)graph" and defers the treatment of general concept graphs to future work ("In the future we plan to investigate this aspect."). The abstract nevertheless claims "a complete explainable system that interprets a set of data ... and describes them in a natural language of choice." Since the Concept2Text input is a tree, the system as presented cannot narrate arbitrary concept graphs; the graph-to-tree conversion is a load-bearing component for the "complete system" claim. The authors should either specify and implement this conversion or explicitly scope their contribution to tree-shaped concepts in the abstract and in the contribution list.
  3. [§4, "General concept narration"] The empirical demonstration enumerates 13,000+ English and 3,200+ Italian sentences, but the only reported validation is that the authors "manually reviewed some samples and they all appear well-formed." No criteria are given for judging "well-formed" or "semantics preserved," and no independent human or automated evaluation is reported. To support the quality claims, the paper should provide a structured evaluation: for example, a random sample of generated sentences rated by multiple human judges for grammaticality and meaning preservation relative to the input, or an automated equivalence check against a formal semantics when the input can be mapped to a logical representation.
minor comments (5)
  1. [§1, bullet "Tree Rewriting"] The phrase "from the conceptual to the the syntax level" contains a duplicated "the"; it should read "to the syntax level."
  2. [§4, comparison with LLMs] The comparison with ChatGPT 3.5 and Gemini is reported without a reproducible protocol: the exact prompt, the number of runs, and the generated outputs are not given. If this comparison is kept, it should be described in detail so that a reader can replicate it; otherwise, the claim that the pipeline is superior to LLMs in semantic integrity is not supported.
  3. [§3.1.2 and §4] The full rule set is not included in the paper, only two code snippets. Since the system's correctness and transparency claims depend on these rules, and the paper emphasizes "verifiability," the authors should make the complete set of rules available in an appendix, a supplementary file, or a persistent repository alongside the sentence list.
  4. [§4, "Data2Concept narration"] The mapping from the integer accuracy (0..100) to the adjective scale (e.g., "excellent accuracy") in Table 1 is not specified. Please define this mapping explicitly or cite the prior work in which it is defined.
  5. [§3.1.2, code snippet] The predicate member_non_var is used in the first rule snippet but is not defined in the text; a one-sentence explanation of its purpose would clarify the code for readers.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the semantic-preservation claim rests on unverified handwritten equivalence rules, which is a correctness risk rather than a circular reduction.

full rationale

This is a systems/architecture paper, not a derivation of a first-principles result or a statistically fitted prediction. The Concept2Text pipeline rewrites concept trees using handwritten rules; the assertion that the output variants are semantically equivalent is an assumption embedded in the rule labels (e.g., equiv_concept), not a conclusion drawn from independent data or from a theorem whose premises include the conclusion. The paper's own validation is manual and non-exhaustive ('We locally checked each rewriting rule behaviour' and 'We manually reviewed some samples'), and no formal semantics for the concept-tree language is given, but that is a verification weakness, not circularity. The self-citations to previous work by the same authors ([3,4,5]) provide the Data2Concept module and a prior trivial generator, but the central new contribution—the stratified, modular rewriting system with its six stages—is implemented and demonstrated independently here. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' own prior work to force a choice, no ansatz is smuggled in via citation, and no known empirical result is merely renamed. Accordingly, no circular step can be exhibited by quoting a specific reduction of the claimed result to its inputs, and the honest finding is 'no significant circularity.'

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on hand-crafted rule sets, lexicons, CSP constraints, and a spanning-tree limitation that are not shipped with the paper. The equivalence and grammatical correctness of generated text is assumed from the authors' manual rule design rather than derived or benchmarked externally.

free parameters (3)
  • accuracy-to-adjective mapping
    Data2Concept outputs an integer accuracy 0-100 mapped to a judgment scale of adjectives (e.g., 'excellent accuracy'), thresholds not specified in the paper. Section 4, Data2Concept narration.
  • CSP ordering constraints
    Word-order constraints for inflection are hand-tuned ('developing accurate order constraints requires careful tuning') and determine sentence well-formedness. Section 3.1.3.
  • rule variant choices = at most 4 per rule
    Each rewriting rule's alternative variants are authored by hand; their count and content are free design choices affecting output diversity. Section 4.
assumptions (5)
  • ad hoc to paper A spanning tree of the original concept graph is sufficient for narration
    Section 3 states 'we assume to work with a simple spanning tree of the original (sub)graph', deferring general graph handling to future work; invalid for concepts requiring multi-branch narrative structure.
  • domain assumption The hand-written equivalence rules preserve semantics
    The claim of perfect semantic preservation relies on the author-defined rewrite rules encoding genuine semantic equivalences; no independent measure or external benchmark validates this. Section 3.1.2 and Section 4.
  • domain assumption Common knowledge ontology with is_a relations is available
    Section 3.1.2 relies on 'common_knowledge_isa' to infer that year is a time and interval of time equals period; such knowledge is not shipped with the system.
  • domain assumption The CSP over word partial orders has a solution and yields a well-formed sentence
    Section 3.1.3: 'By combining these partial orders, we can derive the correct total order of words.' No proof of completeness for English or Italian is given.
  • domain assumption Stratified fixpoint rewriting terminates
    The system repeatedly applies rules until a fixpoint is reached; termination depends on the rule sets not introducing unbounded recursion, which is not formally analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data2Concept2Text: An Explainable Multilingual Framework for Data Analysis Narration." pith.science (2026). https://pith.science/paper/5AMZKGXB

@misc{pith2026250209218,
  author       = {Pith},
  title        = {Pith review of: Data2Concept2Text: An Explainable Multilingual Framework for Data Analysis Narration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5AMZKGXB}},
  note         = {Machine review of arXiv:2502.09218}
}
read the original abstract

This paper presents a complete explainable system that interprets a set of data, abstracts the underlying features and describes them in a natural language of choice. The system relies on two crucial stages: (i) identifying emerging properties from data and transforming them into abstract concepts, and (ii) converting these concepts into natural language. Despite the impressive natural language generation capabilities demonstrated by Large Language Models, their statistical nature and the intricacy of their internal mechanism still force us to employ these techniques as black boxes, forgoing trustworthiness. Developing an explainable pipeline for data interpretation would allow facilitating its use in safety-critical environments like processing medical information and allowing non-experts and visually impaired people to access narrated information. To this end, we believe that the fields of knowledge representation and automated reasoning research could present a valid alternative. Expanding on prior research that tackled the first stage (i), we focus on the second stage, named Concept2Text. Being explainable, data translation is easily modeled through logic-based rules, once again emphasizing the role of declarative programming in achieving AI explainability. This paper explores a Prolog/CLP-based rewriting system to interpret concepts-articulated in terms of classes and relations, plus common knowledge-derived from a generic ontology, generating natural language text. Its main features include hierarchical tree rewritings, modular multilingual generation, support for equivalent variants across semantic, grammar, and lexical levels, and a transparent rule-based system. We outline the architecture and demonstrate its flexibility through some examples capable of generating numerous diverse and equivalent rewritings based on the input concept.

Figures

Figures reproduced from arXiv: 2502.09218 by the authors.

Figure 1
Figure 1. Example of an input concept (the Prolog list and the corresponding tree). [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Example of rewriting with concept equivalence. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Input concept for a full sentence be easily incorporated into rules. 4 Results General concept narration We first test the potential and robustness of the Concept2Text system only. We crafted a concept that is depicted in [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Computed stages for the call for papers concept of Figure 3. From top to bottom: Concept [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 13 canonical work pages

  1. [1]

    (2000): Gene ontology: tool for the unification of biology

    Michael Ashburner et al. (2000): Gene ontology: tool for the unification of biology . Nature genetics 25(1), pp. 25–29, doi:10.1038/75556

    doi:10.1038/75556 2000

  2. [2]

    (2020): Explainable Artificial Intelligence (XAI): Concepts, tax- onomies, opportunities and challenges toward responsible AI

    Alejandro Barredo Arrieta et al. (2020): Explainable Artificial Intelligence (XAI): Concepts, tax- onomies, opportunities and challenges toward responsible AI . Inf. Fusion 58, pp. 82–115, doi:10.1016/j.inffus.2019.12.012

    doi:10.1016/j.inffus.2019.12.012 2020

  3. [3]

    In: Proceedings of the Italia Intelligenza Artificiale (Ital-IA 2023), Pisa, Italy, May 29-30, 2023 , CEUR Workshop Proceedings 3486, CEUR-WS.org, pp

    Flavio Bertini, Alessandro Dal Palù, Andrea Formisano, Andrea Pintus, Sara Rainieri & Luana Salvarani (2023): Students’ Careers and AI: a decision-making support system for Academia . In: Proceedings of the Italia Intelligenza Artificiale (Ital-IA 2023), Pisa, Italy, May 29-30, 2023 , CEUR Workshop Proceedings 3486, CEUR-WS.org, pp. 272–277. Available at ...

    work page 2023

  4. [4]

    Alessandro Dal Palù, Agostino Dovier & Andrea Formisano (2023): Towards Explainable Data-to-text Gen- eration. In Agostino Dovier & Andrea Formisano, editors: Proceedings of the 38th Italian Conference on Computational Logic, Udine, Italy, June 21-23, 2023 , CEUR Workshop Proceedings 3428, CEUR-WS.org, pp. 1–15. Available at https://ceur-ws.org/Vol-3428/ ...

    work page 2023

  5. [5]

    In Enrico Pontelli et al., editors: Proc

    Alessandro Dal Palù, Agostino Dovier & Andrea Formisano (2023): An xAI Approach for Data-to-Text Pro- cessing with ASP. In Enrico Pontelli et al., editors: Proc. 39th International Conference on Logic Program- ming, ICLP 2023, Imperial College London, UK, 9th July 2023 - 15th July 2023, EPTCS 385, pp. 353–366, doi:10.4204/EPTCS.385.38

    work page doi:10.4204/eptcs.385.38 2023

  6. [6]

    Cambridge Univ

    Marcel den Dikken (2013): The Cambridge handbook of generative syntax . Cambridge Univ. Press, doi:10.1017/CBO9780511804571

    work page doi:10.1017/cbo9780511804571 2013

  7. [7]

    Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act)

    European Commission (2021): Regulation of the European Parliament and of the Council. Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) . https://eur-lex.europa.eu/ legal-content/EN/TXT/HTML/?uri=CELEX:52021PC0206

    work page 2021

  8. [8]

    New England Journal of Medicine 379(15), pp

    Melissa A Haendel, Christopher G Chute & Peter N Robinson (2018): Classification, ontology, and precision medicine. New England Journal of Medicine 379(15), pp. 1452–1462, doi:10.1056/NEJMra1615014

    doi:10.1056/nejmra1615014 2018

Show all 22 references
  1. [9]

    (2023): Survey of hallucination in natural language generation

    Ziwei Ji et al. (2023): Survey of hallucination in natural language generation . ACM Computing Surveys 55(12), pp. 1–38, doi:10.48550/ARXIV .2405.20315

    2023 doi arXiv

  2. [10]

    (2021): The human phenotype ontology in 2021

    Sebastian Köhler et al. (2021): The human phenotype ontology in 2021. Nucleic acids research 49(D1), pp. D1207–D1217, doi:10.1093/NAR/GKAA1043

    2021 doi

  3. [11]

    Yu Lin & Yongqun He (2012): Ontology representation and analysis of vaccine formulation and administra- tion and their effects on vaccine immune responses. J. Biomed. Semant.3, pp. 1–15, doi:10.1186/2041-1480- 3-17

    2012 doi

  4. [12]

    New Generation Computing 1(2), pp

    Yuji Matsumoto, Hozumi Tanaka, Hideki Hirakawa, Hideo Miyoshi & Hideki Yasukawa (1983): BUP: a bottom-up parser embedded in Prolog . New Generation Computing 1(2), pp. 145–158, doi:10.1007/BF03037421

    1983 doi

  5. [13]

    McGuinness, Frank Van Harmelen et al

    Deborah L. McGuinness, Frank Van Harmelen et al. (Updated in 2009): OWL web ontology language overview. W3C recommendation 10 February 2004. Available at www.w3.org/TR/owl-features/

    2009

  6. [14]

    Noy, Mark A

    Natalya F. Noy, Mark A. Musen, José L.V . Mejino Jr & Cornelius Rosse (2004): Pushing the envelope: challenges in a frame-based representation of human anatomy . Data Knowl. Eng. 48(3), pp. 335–359, doi:10.1016/J.DATAK.2003.06.002

    2004 doi

  7. [15]

    Fernando C. N. Pereira (1987): Grammars and Logics of Partial Information. In Jean-Louis Lassez, editor: Logic Programming, Proceedings of the Fourth International Conference, Melbourne, Victoria, Australia, May 25-29, 1987, MIT Press, pp. 989–1013. ISBN:0262121255

    1987

  8. [16]

    Fernando C. N. Pereira & Stuart M. Shieber (2002): Prolog and natural-language analysis . Microtome Publishing. ISBN:0937073180

    2002

  9. [17]

    Schriml et al

    Lynn M. Schriml et al. (2019): Human Disease Ontology 2018 update: classification, content and workflow expansion. Nucleic acids research 47(D1), pp. D955–D962, doi:10.1093/NAR/GKY1032

    2019 doi

  10. [18]

    (2007): The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration

    Barry Smith et al. (2007): The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration. Nature biotechnology 25(11), pp. 1251–1255, doi:10.1038/nbt1346

    2007 doi

  11. [19]

    In: Proceedings of the 31st AAAI conference on artificial intelligence , pp

    Robyn Speer, Joshua Chin & Catherine Havasi (2017): Conceptnet 5.5: An open multilingual graph of general knowledge. In: Proceedings of the 31st AAAI conference on artificial intelligence , pp. 4444–4451, doi:10.1609/AAAI.V31I1.11164

    2017 doi

  12. [20]

    International Handbooks on Infor- mation Systems, Springer, doi:10.1007/978-3-540-92673-3

    Steffen Staab & Rudi Studer, editors (2009): Handbook on Ontologies. International Handbooks on Infor- mation Systems, Springer, doi:10.1007/978-3-540-92673-3

    2009 doi

  13. [21]

    arXiv preprint arXiv:2307.12966, doi:10.48550/ARXIV .2307.12966

    Yufei Wang, Wanjun Zhong, Liangyou Li, Fei Mi, Xingshan Zeng, Wenyong Huang, Lifeng Shang, Xin Jiang & Qun Liu (2023): Aligning large language models with human: A survey. arXiv preprint arXiv:2307.12966, doi:10.48550/ARXIV .2307.12966

    2023 doi arXiv

  14. [22]

    (2023): A survey of large language models

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Be- ichen Zhang, Junjie Zhang, Zican Dong et al. (2023): A survey of large language models . arXiv preprint arXiv:2303.18223, doi:10.48550/ARXIV .2303.18223. 152 Data2Concept2Text: Explain...

    2023 doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.