A Taxonomy for Design and Evaluation of Prompt-Based Natural Language Explanations

Isar Nejadgholi; Jonathan Boisvert; Marc-Antoine Drouin; Mona Omidyeganeh

REVIEW 3 major objections 5 minor 1 cited by

A Taxonomy for Design and Evaluation of Prompt-Based Natural Language Explanations

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a three-part taxonomy with fifteen evaluation properties as a checklist for designing and assessing prompt-based natural language explanations.

desk verdict A well-scoped, useful taxonomy for prompt-based NLEs, but its completeness rests on an undocumented expert consensus; it deserves review, not blind adoption. read the letter →

arxiv 2507.10585 v1 pith:63ZOTRQ5 submitted 2025-07-11 cs.CL cs.AI

Isar Nejadgholi , Mona Omidyeganeh , Marc-Antoine Drouin , Jonathan Boisvert This is my paper · ORCID

classification cs.CLcs.AI

keywords naturallanguageexplanationsXAItaxonomyprompt-basedexplainableAIgovernanceevaluationpropertieslargemodelstransparency

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

This paper argues that existing explainable-AI taxonomies do not capture what is special about explanations written in natural language, and proposes a dedicated taxonomy for prompt-based natural language explanations (NLEs). The taxonomy organizes NLE design and evaluation into three dimensions: context (task, data, audience, goal), generation and presentation (model type, prompt input, interactivity, output type, presentation form), and evaluation (content, presentation, and user-centered properties plus evaluation setting). It distills fifteen desirable properties, including correctness, comprehensibility, translucence, actionability, and novelty. If adopted, the taxonomy gives researchers, auditors, and policymakers a common checklist for designing, comparing, and verifying the explanations AI systems generate in plain language.

What carries the argument

The load-bearing object is the two-table taxonomy: Table 1 maps context and generation/presentation, and Table 2 lists the fifteen evaluation properties. The machinery works by forcing each explanation design to be described on three axes—context, generation/presentation, evaluation—with each axis decomposed into subcategories; this decomposition turns a vague notion like 'a good explanation' into a checklist that can be applied before and after building a prompt.

What would settle it

A concrete test would be a coding study in which two independent teams of experts each classify a diverse sample of real prompt-based NLEs, from domains such as healthcare and finance, into the taxonomy's subcategories; if inter-rater agreement is low, or a substantial fraction of explanations cannot be assigned to any subcategory, the taxonomy's categories are not comprehensive or distinct. A second falsifier is a user study in which explanations designed with the taxonomy do no better on the fifteen properties than untailored explanations in a head-to-head comparison.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that the standard XAI taxonomy can be adapted into a structured, model-agnostic checklist specific to post-hoc, model-agnostic, local, prompt-based natural language explanations. The adaptation adds an Audience subcategory with five roles (creator, operator, executor, decision subject, examiner) and an Explanation Goal subcategory, reworks the explanation-generation module around prompt inputs and dialogue-based interactivity, and extends the twelve-property evaluation framework to fifteen properties for NLEs. The paper presents this as a practical instrument for technical AI governance: it lets stakeholders specify what an explanation is for, generate it in a form suited to that purpose, and evaluate it along properties that can be tested in functionally grounded, human-grounded, or application-grounded settings.

Load-bearing premise

The taxonomy is only as sound as the two source frameworks it adapts—the broad XAI meta-taxonomy and the Co-12 property list—and the informal expert consensus process that chose the adaptations; if either source missed important facets of natural language explanations, or the expert panel was unrepresentative, the taxonomy inherits those gaps.

Editorial extensions

If this is right

Designers can use the checklist to construct system prompts that specify task type, data type, audience, explanation goal, output type, and presentation format, as demonstrated in the traffic-anomaly use case.
Evaluators can choose among functionally grounded, human-grounded, and application-grounded settings for each of the fifteen properties, making evaluation plans more explicit and comparable across studies.
Auditors and policymakers gain a shared vocabulary for requesting and reviewing natural language explanations, supporting transparency and accountability requirements.
The taxonomy highlights tensions between properties, such as perfect correctness reducing comprehensibility, prompting explicit trade-off decisions rather than implicit ones.
Researchers can position new NLE work within the taxonomy and identify which properties a proposed method targets.

Reading between the lines

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

A next step the paper does not take is to turn the taxonomy into a scoring rubric and measure inter-rater reliability across independent annotators; that would test whether the categories are stable in practice.
The taxonomy could be operationalized as a multi-objective optimization for prompt search, where the fifteen properties serve as objectives and conflicting pairs, such as correctness versus comprehensibility, define a Pareto front.
Because the evaluation properties are defined independently of any particular model, the same checklist could plausibly be applied beyond prompt-based LLMs to retrieval-augmented or fine-tuned explanation generators with minor adjustments.
The audience-role subcategory suggests testable hypotheses, for example that examiner-oriented explanations should emphasize traceability and policy alignment, and user studies could verify that role-specific prompts improve task performance.

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.

Desk Editor's Note

read the letter

Bottom line: this is a useful, clearly scoped taxonomy proposal for prompt-based NLEs, not a demonstrated standard. The new subcategories (Audience, Explanation Goal, Translucence, Novelty, Personalization/Actionability) are genuinely missing from the source XAI taxonomies, and the paper's explicit scope (post-hoc, model-agnostic, local) keeps the claims contained. I read it as a checklist for designers and evaluators, and it does that job.

What it does well: The adaptation is grounded in the right prior work, and the paper is transparent about what it omitted and why. The example in Appendix B shows the taxonomy in action with a real system prompt and response, which makes the abstract categories concrete. The discussion of trade-offs between properties, especially faithfulness vs comprehensibility, is honest and doesn't pretend the framework resolves them.

Where it's soft: The validity of the taxonomy depends on two things the paper doesn't demonstrate: the completeness of the source taxonomies (Schwalbe & Finzel, Co-12) and the quality of the expert consensus process. We're told only that a small group of experts, including the authors, iteratively adapted the framework. No inter-rater reliability, no independence check, no external validation. The paper says validation is future work, which is fine for a workshop paper, but it should be explicit that the taxonomy is a hypothesis, not a finding. The overlap between Interactivity and Input is a bit blurry, and some user-centered properties (Coherence vs Personalization) could be hard to distinguish in practice. These are minor.

The citation pattern looks fine. It uses the sources it builds on and doesn't bury prior work.

Who it's for: people building or evaluating NLE systems, and anyone in technical governance who needs a shared vocabulary. A serious referee could push the authors to document the expert selection process, add a comparison to other NLE-specific taxonomies, and clarify category boundaries. I'd send it to review.

Referee Report

3 major / 5 minor

Summary. The paper proposes a taxonomy for prompt-based natural language explanations (NLEs), adapted from the Schwalbe and Finzel XAI meta-taxonomy and the Nauta et al. Co-12 property framework. The taxonomy is structured in three dimensions: Context Definition, Generation and Presentation, and Evaluation, with 15 evaluation properties. The authors scope the work to post-hoc, model-agnostic, local prompt-based NLEs and illustrate the taxonomy with a use case in airborne anomaly detection, including a worked prompt and model response. The paper is explicitly a proposal: validation is deferred to future work, and several limitations (faithfulness, subjectivity, trade-offs) are acknowledged in Section 5.

Significance. If the taxonomy is accepted, it gives researchers, auditors, and policymakers a structured checklist for characterizing and evaluating prompt-based NLEs, filling a genuine gap since existing XAI taxonomies are not NLE-specific. The paper builds transparently on established sources (Schwalbe and Finzel, Nauta et al., Liao et al.), clearly scopes its claims, and includes a concrete example that helps operationalize the abstraction. Its limitations are honestly stated, including the lack of empirical validation and the acknowledged difficulty of faithfulness. As a proposal, the contribution is useful, but the present form leaves several load-bearing validity questions unanswered.

major comments (3)

[Section 2 and Section 4/Appendix B] The example use case in Section 4 and Appendix B does not satisfy the scope defined in Section 2. The paper restricts itself to post-hoc, model-agnostic, and local prompt-based NLEs, but in the anomaly-detection example the same vision-language model both detects the anomaly and generates the explanation (the system prompt in Figure B.1 instructs the model to 'detect and explain anomalies'), which is ante-hoc and model-specific rather than post-hoc and model-agnostic. This inconsistency undermines the claim that the taxonomy applies to the stated scope. Please either revise the example to use a separate post-hoc explainer or explicitly extend the scope to include same-model explanations and adjust the definitions in Section 2 accordingly.
[Section 3.3 and Table 2] The adaptation of the Co-12 framework conflates two distinct constructs. In Nauta et al., 'Covariate Complexity' refers to the number of covariates or features used in the explanation, whereas Table 2 defines 'Comprehensibility' as 'Uses human-understandable concepts and relations.' These are not equivalent, and renaming one to the other changes the meaning of the property. Please either retain the original construct with a clearer justification for the rename or define Comprehensibility as a new property and describe what happens to Covariate Complexity (e.g., whether it is omitted or folded into another property).
[Section 2 (Methodological approach)] The expert consensus process is described only as 'an iterative, consensus-driven process involving experts from multiple domains,' with no information on the number of experts, their selection criteria, their independence from the authors, or any inter-rater reliability assessment. Because the taxonomy's completeness and the specific adaptations are the central claims of the paper, this opaque methodology weakens confidence in the result. Please report the panel details and reliability metrics, or alternatively reframe the contribution as a proposed taxonomy whose validation is explicitly left to future work, softening the 'provides a framework' claim in the abstract and conclusion.

minor comments (5)

[Section 3.1] There is a typo: 'meat-taxonomy' should be 'meta-taxonomy' in the first paragraph of Section 3.1.
[Section 3.1] The phrase 'we adopt the golas of NLE generation' contains a typo; 'golas' should be 'goals'.
[Section 3.2 and Table 1] The relationship between Interactivity and Input is acknowledged as blurred in the text, but Table 1 lists them as separate subcategories. Consider clarifying the boundary or merging them into a single subcategory, since the current presentation leaves the reader unsure how to apply them distinctly.
[Appendix B, Figure B.1] In the sample response, 'Anomaly Detected: Yes' uses a capital 'Yes' while the prompt specifies the output should be 'YES' or 'NO' (all caps). Please ensure consistency between the prompt specification and the example response.
[Table 2] The 'Confidence' property in the Presentation category is described as 'Communicates the model's certainty or uncertainty.' It would be clearer to specify whether this refers to the confidence in the prediction or in the explanation itself, since these can differ substantially for NLEs.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the taxonomy is a transparent synthesis of external prior taxonomies, with empirical validation explicitly deferred.

full rationale

This paper makes no empirical prediction and contains no fitted parameters. Its contribution is a proposed taxonomy assembled from external prior taxonomies: Schwalbe & Finzel (2024) supplies the three-component structure, Nauta et al. (2023) supplies the Co-12 evaluation properties, Liao et al. (2022) informs the usage-context evaluation framing, and Tomsett et al. (2018) and Chen et al. (2022) ground the Audience and Explanation Goal subcategories. Each adaptation is explicit rather than hidden: the paper states 'we rename Covariate Complexity as Comprehensibility', 'we decompose the original Context property', 'we add Translucence', and 'we include Novelty'. The taxonomy is not used to justify itself: Section 6 explicitly defers empirical validation ('Future work involves validating its use through experiments in collaboration with human-computer interaction experts'). The expert consensus process described in Section 2 is a methodological limitation (a small panel that includes the authors, with no reported inter-rater reliability), but it is not a circular derivation because the resulting categories are not presupposed by any claim the paper purports to derive from them. There are no equations, no fitted inputs renamed as predictions, and no load-bearing self-citation chain; every cited source is external prior work. The central claim that the taxonomy 'provides a framework' is an unsupported proposal pending validation, but unsupportedness is a correctness-risk concern, not circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters or invented entities. The central claim rests on the completeness of the cited XAI taxonomies and on expert consensus as a validation method. The axioms below list the background assumptions from prior literature and scoping choices.

assumptions (4)

domain assumption The Schwalbe and Finzel (2024) XAI meta-taxonomy is a comprehensive and valid synthesis of XAI concepts.
Section 2: 'Our work builds on the XAI concepts synthesized in a recent meta-analysis by Schwalbe & Finzel (2024)'. The entire adaptation starts from this taxonomy; if it were incomplete, the proposed taxonomy would inherit gaps.
domain assumption The Co-12 explanation properties (Nauta et al., 2023) are an appropriate basis for NLE evaluation properties.
Section 3.3: 'we start with a framework with twelve explanation properties, Co-12, originally proposed by Nauta et al. (2023)'. The 15 properties in Table 2 are derived from Co-12 for all additions.
domain assumption An iterative consensus-driven process with a small set of experts suffices to validate which components of the meta-taxonomy apply to NLEs.
Section 2: 'informed by an iterative, consensus-driven process involving experts from multiple domains, including the authors'. No inter-rater reliability or empirical test is reported.
domain assumption Limiting scope to post-hoc, model-agnostic, local prompt-based NLEs is a meaningful and non-exclusionary assumption.
Section 2 narrows scope; the taxonomy does not cover inherently interpretable models, global explanations, or ante-hoc methods. It assumes this subset captures the main governance needs.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Taxonomy for Design and Evaluation of Prompt-Based Natural Language Explanations." pith.science (2026). https://pith.science/paper/63ZOTRQ5

@misc{pith2026250710585,
  author       = {Pith},
  title        = {Pith review of: A Taxonomy for Design and Evaluation of Prompt-Based Natural Language Explanations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/63ZOTRQ5}},
  note         = {Machine review of arXiv:2507.10585}
}

read the original abstract

Effective AI governance requires structured approaches for stakeholders to access and verify AI system behavior. With the rise of large language models, Natural Language Explanations (NLEs) are now key to articulating model behavior, which necessitates a focused examination of their characteristics and governance implications. We draw on Explainable AI (XAI) literature to create an updated XAI taxonomy, adapted to prompt-based NLEs, across three dimensions: (1) Context, including task, data, audience, and goals; (2) Generation and Presentation, covering generation methods, inputs, interactivity, outputs, and forms; and (3) Evaluation, focusing on content, presentation, and user-centered properties, as well as the setting of the evaluation. This taxonomy provides a framework for researchers, auditors, and policymakers to characterize, design, and enhance NLEs for transparent AI systems.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Traccia: An OpenTelemetry-Based Governance Platform for AI Systems
cs.AI 2026-07 conditional novelty 5.0 of 10

Traccia is a seven-layer OpenTelemetry-based pipeline that converts AI execution traces into hash-protected, regulation-mapped compliance evidence for EU AI Act audits.

Reference graph

Works this paper leans on

16 extracted references · 8 canonical work pages · cited by 1 Pith paper

[4]

and Kim, B

Doshi-Velez, F. and Kim, B. Towards a rigorous sci- ence of interpretable machine learning. arXiv preprint arXiv:1702.08608,

arXiv
[6]

doi: 10.18653/v1/2020.acl-main

Association for Compu- tational Linguistics. doi: 10.18653/v1/2020.acl-main

doi:10.18653/v1/2020.acl-main 2020
[9]

G., Hofman, J

Poursabzi-Sangdeh, F., Goldstein, D. G., Hofman, J. M., Wortman Vaughan, J. W., and Wallach, H. Manipulating and measuring model interpretability. In Proceedings of the 2021 CHI conference on human factors in computing systems, pp. 1–52,

work page 2021
[10]

Open problems in technical ai governance

Reuel, A., Bucknall, B., Casper, S., Fist, T., Soder, L., Aarne, O., Hammond, L., Ibrahim, L., Chan, A., Wills, P., et al. Open problems in technical ai governance. arXiv preprint arXiv:2407.14981,

arXiv
[11]

A review of taxonomies of explainable artifi- cial intelligence (xai) methods

Speith, T. A review of taxonomies of explainable artifi- cial intelligence (xai) methods. In Proceedings of the 2022 ACM conference on fairness, accountability, and transparency, pp. 2239–2250,

work page 2022
[15]

V ., and Bellamy, R

Zhang, Y ., Liao, Q. V ., and Bellamy, R. K. Effect of confi- dence and explanation on accuracy and trust calibration in ai-assisted decision making. In Proceedings of the 2020 conference on fairness, accountability, and transparency, pp. 295–305,

work page 2020
[16]

Meta-Taxonomy of XAI Figure A.1 presents the meta-taxonomy, whose components we systematically analyzed to develop our proposed taxonomy

8 A Design and Evaluation Taxonomy for Natural Language Explanations A. Meta-Taxonomy of XAI Figure A.1 presents the meta-taxonomy, whose components we systematically analyzed to develop our proposed taxonomy. Figure A.1. The XAI meta-taxonomy proposed by Schwalbe & Finzel (2024) (Page 50, Figure 7). 9 A Design and Evaluation Taxonomy for Natural Language...

work page 2024
[386]

acl-main.386/

URL https://aclanthology.org/2020. acl-main.386/. Kayser, M. G., Menzat, B., Emde, C., Bercean, B. A., Novak, A., Morgado, A. T. E., Papiez, B., Gaube, S., Lukasiewicz, T., and Camburu, O.-M. Fool me once? contrasting textual and visual explanations in a clinical decision-support setting. In Proceedings of the 2024 Conference on Empirical Methods in Natur...

work page 2020

Show all 16 references

[2017]

URL https://ojs.aaai.org/aimagazine/index

doi: 10.1609/aimag.v38i3.2741. URL https://ojs.aaai.org/aimagazine/index. php/aimagazine/article/view/2741. Gu, J., Jiang, X., Shi, Z., Tan, H., Zhai, X., Xu, C., Li, W., Shen, Y ., Ma, S., Liu, H., et al. A survey on llm-as-a- judge. arXiv preprint arXiv:2411.15594,

arXiv
[2018]

and Rad, P

Das, A. and Rad, P. Opportunities and challenges in explain- able artificial intelligence (xai): A survey. arXiv preprint arXiv:2006.11371,

2006 arXiv
[2020]

T., and Weld, D

Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., and Weld, D. Does the whole exceed its parts? the effect of ai explanations on complementary team performance. In Proceedings of the 2021 CHI con- ference on human factors in computing systems, pp. 1–16,

2021
[2021]

Reframing human-AI collaboration for generat- ing free-text explanations

Wiegreffe, S., Hessel, J., Swayamdipta, S., Riedl, M., and Choi, Y . Reframing human-AI collaboration for generat- ing free-text explanations. In Carpuat, M., de Marneffe, M.-C., and Meza Ruiz, I. V . (eds.), Proceedings of the 2022 Conference of the North American Chapter of ...

2022
[2022]

doi: 10.18653/v1/2022.naacl-main.47

Association for Computational Linguistics. doi: 10.18653/v1/2022.naacl-main.47. URL https:// aclanthology.org/2022.naacl-main.47/. Ye, X. and Durrett, G. The unreliability of explanations in few-shot prompting for textual reasoning. Advances in neural information processing sy...

2022 doi
[2023]

Machine explanations and human understanding

Chen, C., Feng, S., Sharma, A., and Tan, C. Machine explanations and human understanding. arXiv preprint arXiv:2202.04092,

arXiv
[2024]

Meyer, L. S. and Zhu, J. Slide to explore’what if’: An analysis of explainable interfaces. In Adjunct Proceed- ings of the 2024 Nordic Conference on Human-Computer Interaction, pp. 1–6,

2024
[2025]

Interpretable to whom? a role-based model for analyzing interpretable machine learning sys- tems

Tomsett, R., Braines, D., Harborne, D., Preece, A., and Chakraborty, S. Interpretable to whom? a role-based model for analyzing interpretable machine learning sys- tems. arXiv preprint arXiv:1806.07552,

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [4]

and Kim, B

Doshi-Velez, F. and Kim, B. Towards a rigorous sci- ence of interpretable machine learning. arXiv preprint arXiv:1702.08608,

arXiv

[2] [6]

doi: 10.18653/v1/2020.acl-main

Association for Compu- tational Linguistics. doi: 10.18653/v1/2020.acl-main

doi:10.18653/v1/2020.acl-main 2020

[3] [9]

G., Hofman, J

Poursabzi-Sangdeh, F., Goldstein, D. G., Hofman, J. M., Wortman Vaughan, J. W., and Wallach, H. Manipulating and measuring model interpretability. In Proceedings of the 2021 CHI conference on human factors in computing systems, pp. 1–52,

work page 2021

[4] [10]

Open problems in technical ai governance

Reuel, A., Bucknall, B., Casper, S., Fist, T., Soder, L., Aarne, O., Hammond, L., Ibrahim, L., Chan, A., Wills, P., et al. Open problems in technical ai governance. arXiv preprint arXiv:2407.14981,

arXiv

[5] [11]

A review of taxonomies of explainable artifi- cial intelligence (xai) methods

Speith, T. A review of taxonomies of explainable artifi- cial intelligence (xai) methods. In Proceedings of the 2022 ACM conference on fairness, accountability, and transparency, pp. 2239–2250,

work page 2022

[6] [15]

V ., and Bellamy, R

Zhang, Y ., Liao, Q. V ., and Bellamy, R. K. Effect of confi- dence and explanation on accuracy and trust calibration in ai-assisted decision making. In Proceedings of the 2020 conference on fairness, accountability, and transparency, pp. 295–305,

work page 2020

[7] [16]

Meta-Taxonomy of XAI Figure A.1 presents the meta-taxonomy, whose components we systematically analyzed to develop our proposed taxonomy

8 A Design and Evaluation Taxonomy for Natural Language Explanations A. Meta-Taxonomy of XAI Figure A.1 presents the meta-taxonomy, whose components we systematically analyzed to develop our proposed taxonomy. Figure A.1. The XAI meta-taxonomy proposed by Schwalbe & Finzel (2024) (Page 50, Figure 7). 9 A Design and Evaluation Taxonomy for Natural Language...

work page 2024

[8] [386]

acl-main.386/

URL https://aclanthology.org/2020. acl-main.386/. Kayser, M. G., Menzat, B., Emde, C., Bercean, B. A., Novak, A., Morgado, A. T. E., Papiez, B., Gaube, S., Lukasiewicz, T., and Camburu, O.-M. Fool me once? contrasting textual and visual explanations in a clinical decision-support setting. In Proceedings of the 2024 Conference on Empirical Methods in Natur...

work page 2020