Pith. sign in

REVIEW 5 major objections 5 minor 38 references

MetaExplainer: A Framework to Generate Multi-Type User-Centered Explanations for AI Systems

T0 review · 5 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 MetaExplainer claims a three-stage pipeline can turn user questions into diverse, natural-language AI explanations.

desk verdict A genuinely modular pipeline for multi-type explanations, but the end-to-end claim is undercut by the Decompose stage's 0.00 F1 on Contrastive questions. read the letter →

arxiv 2508.00300 v2 pith:JJGCGV6A submitted 2025-08-01 cs.HC cs.AIcs.LG

Shruthi Chari , Oshani Seneviratne , Prithwish Chakraborty , Pablo Meyer , Deborah L. McGuinness This is my paper
classification cs.HCcs.AIcs.LG
keywords explainableAIuser-centeredexplanationsExplanationOntologylargelanguagemodelsretrieval-augmentedgenerationcounterfactualcontrastivehealthcare
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 tries to close the gap between the raw outputs of model explainers and the questions real users actually ask, by introducing MetaExplainer, a three-stage neuro-symbolic framework. The framework decomposes a user's natural-language question into a machine-readable form, delegates the question to established explainer methods, and synthesizes the results into natural-language explanations. The paper argues that an Explanation Ontology—a structured catalog of explanation types and their links to explainer methods—is what lets the system choose the right explanation type and method for a given question. Results on a diabetes prediction task report per-stage scores (59.06% F1 for machine interpretation, 0.71 faithfulness for feature-based explanations, 67% context utilization in synthesis) and a user study in which over 90% of participants responded positively to trust and curiosity questions. The broader claim is that a modular, ontology-guided design can make user-centered explainability practical across domains.

What carries the argument

The load-bearing mechanism is the Explanation Ontology (EO), a structured representation that says which explanation types exist, which explainer methods can produce them, and what output modalities and templates those methods use. Around the EO, the decompose-delegate-synthesis pipeline works as follows: a fine-tuned Llama model turns a user question into a predicate-logic machine interpretation that names the explanation type, action, and feature filters; SPARQL queries against the EO select explainer methods such as SHAP, DiCE, ProtoDash, and RuleXAI; and a RAG step, implemented with LlamaIndex's PandasQueryEngine, aligns the explainer outputs to the type's template and writes the final natural-language explanation.

What would settle it

Take a held-out set of user questions about a clinical model, have two annotators independently label the intended explanation type for each question, run MetaExplainer, and compare its chosen explanation type against the annotators. If agreement is no better than chance, or if replacing the ontology with a simple keyword-based mapper yields equal or better user-rated explanations, the central claim that ontology-guided selection is what makes the framework user-centered would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that user-centered explanations can be generated by chaining a fine-tuned LLM parser, ontology-driven explainer selection, and retrieval-augmented natural-language generation, and that the chain is measurable stage-by-stage. The paper demonstrates this on a tabular diabetes dataset with five explanation types—contrastive, counterfactual, rationale, case-based, and data—and reports the strongest support from the machine-interpretation F1 of 59.06%, explainer faithfulness of 0.71, context-utilization of 67%, and broadly positive user responses. The intended contribution is not a new single explainer but a reusable architecture that selects among existing explainers according to an explicit symbolic model of explanation types.

Load-bearing premise

The framework assumes the Explanation Ontology correctly and completely maps user-question explanation types to the explainer methods that should answer them, so if that mapping is wrong or incomplete the system will ask the wrong explainer and the rest of the chain cannot fix it.

Editorial extensions

If this is right

  • A system developer can add a new explanation type or explainer method by updating the Explanation Ontology rather than rewriting the entire framework.
  • The same three-stage chain can in principle be pointed at another dataset by retraining or re-prompting the decomposition model and registering the data's feature schema, since the explainer selection is ontology-driven.
  • Each stage emits an inspectable intermediate output, giving a trace from user question to explanation that can localize where a bad explanation originates.
  • If the framework works as reported, non-expert users can interrogate a model with questions phrased as 'why,' 'why not,' 'what if,' and 'what cases' and receive answers in familiar language instead of raw feature-importance plots or rules.

Reading between the lines

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

  • The paper tests only a small tabular dataset, so the most direct next test is whether the Decompose stage and ontology mappings survive transfer to text, image, or time-series data without substantial per-domain engineering.
  • User responses show curiosity is satisfied but confidence in the system is low, which suggests the framework itself needs an explanation of its own explainer choices before users will rely on it for decisions.
  • The ontology mappings are treated as ground truth here; converting them into a benchmark of independently annotated question–explanation-type pairs would allow the central assumption to be tested directly.
  • A natural ablation would remove the EO and replace it with a keyword-based or LLM-only mapper; if that ablation matches the reported metrics, the ontology's specific contribution would need to be re-argued.
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

5 major / 5 minor

Summary. The paper introduces MetaExplainer, a three-stage neuro-symbolic framework for generating user-centered explanations in response to natural-language questions. The Decompose stage uses a fine-tuned Llama3 model to parse questions into an explanation type, a predicate-logic machine interpretation, an action, and a likelihood field; the Delegate stage selects and runs model explainers based on Explanation Ontology (EO) mappings; the Synthesis stage uses RAG with EO templates to produce natural-language explanations. The authors report per-stage metrics (machine-interpretation F1 59.06%, Delegate faithfulness 0.71, Synthesis context-utilization 67%) and a user study with 20 participants. The paper claims support for five explanation types: Contrastive, Counterfactual, Rationale, Case-Based, and Data.

Significance. If fully supported, the framework would be a useful modular contribution to interactive explainability, combining LLMs, ontology-guided routing, and model explainers. The paper's strengths include a clearly described architecture, an open-source code release, and a per-stage evaluation using community metrics. However, the current evidence does not fully support the central claim of multi-type user-centered explanation generation, because the Decompose model has zero F1 for Contrastive identification and synthesis faithfulness is only 25%.

major comments (5)
  1. [Appendix D.2, Table 7] The fine-tuned Llama3 model used in Decompose achieves 0.00 precision, recall, and F1 for Contrastive Explanation identification (support = 5) and only 0.12 F1 for Case-Based Explanation. Since §3.1 states that Delegate selects explainers based on the predicted explanation type, the system cannot autonomously handle 'why not' questions, which is one of the five explanation types claimed in the abstract and §5. If the Contrastive explanations in the user study (Table 12) were generated with oracle type labels, this must be stated; otherwise the results are inconsistent. The authors should report per-type end-to-end performance or retrain/improve the Decompose model.
  2. [Table 3] Synthesis faithfulness is 25% under the RAGAS metric. This means the generated natural-language explanations are faithful to the retrieved context only a quarter of the time. The abstract and §4.2 emphasize 'high performance across all stages' and cite context-utilization (67%) but omit this low faithfulness value. Since faithfulness is central to trustworthy explanations, the authors should analyze the causes and discuss the implications for the framework's claim of generating user-centered explanations.
  3. [§2.3, §4.2] There is no experimental comparison with MegaExplainer, the closest related work, or with a direct LLM baseline that generates explanations without the decompose-delegate-synthesis pipeline. The claim that MetaExplainer is more general and flexible than MegaExplainer is not empirically demonstrated. An ablation removing the EO templates or using a single prompted LLM to produce the final explanation would help determine whether the neuro-symbolic architecture provides measurable benefit.
  4. [§3.3, Tables 4 and 12] The user study has only 20 AI-literate participants, not the target clinical users, and no confidence intervals or significance tests are reported. The abstract's statement that 'over 90% of user-study participants responded positively to trust and curiosity questions' is misleading: Table 4 shows overall positive respondents at 67.41% and only 33.33% agreement with 'I am confident in the MetaExplainer.' The high percentages are for curiosity items (e.g., 93.33% for 'I want to know more'). The authors should present the full breakdown and temper the summary in the abstract.
  5. [§2.2, §3.1] The framework's correctness depends on the Explanation Ontology (EO) mappings from explanation types to explainer methods and templates, which were created by the authors (Chari et al. 2023) and are not independently validated in this paper. If a mapping is wrong, Delegate will run the wrong explainer and Synthesis will use the wrong template, breaking the user-centered explanation chain. The authors should provide evidence for the validity of the EO mappings (e.g., expert validation, ablation, or comparison with an alternative mapping).
minor comments (5)
  1. [§3.3] The phrase 'a 02 scale' should be 'a 0-2 scale.'
  2. [§4.2] The reference to 'the Appendix (Sec. 9)' appears to point to the wrong section; the per-type identification scores are in Appendix D.2.
  3. [§5] The sentence 'users are curious and eager to use our MetaExplainer (Tab. 12)' refers to a satisfaction table; the curiosity/eagerness data are in Table 4. The reference should be corrected.
  4. [Table 2] The metrics in Table 2 are presented as aggregate means without clearly indicating which value corresponds to which explainer method (e.g., Protodash vs. DiCE), making it hard to interpret the differences.
  5. [Appendix D.2] The word 'consoldiated' should be 'consolidated.'

Circularity Check

2 steps flagged · score 6.0 of 10

Contrastive-type user-study results reduce to oracle labels, not Decompose predictions; explainer selection relies on the authors' own Explanation Ontology without independent validation.

  1. fitted input called prediction [Appendix D.2 Table 7 vs Appendix D.5 Table 12; Sec. 3.1 Decompose/Delegate]
    "Table 7: 'Contrastive Explanation 0.00 0.00 0.00 5' ... Table 12: 'MetaExplainer performance for satisfaction as rated by users. The first row shows the percentage of users with positive (>=Neutral ...) assessment ... Predicted Explanation Type ... Contrastive ... 90.48'"

    The paper's multi-type claim requires Decompose to route each user question to the right explanation type, and Delegate uses that type to select explainers and templates. Table 7 reports that the fine-tuned Llama3 Decompose model achieves 0.00 precision, 0.00 recall, and 0.00 F1 for Contrastive questions on the 53-instance test set. Therefore the Contrastive explanations evaluated in Table 12 could not have been generated through the measured Decompose prediction path; either their type label was supplied as an oracle/gold input to Delegate and Synthesis, or the explanations were generated using a different, unreported selector.

  2. self citation load bearing [Sec. 2.2 and Sec. 3.1 Delegate]
    "In the Explanation Ontology (EO) Chari et al. (May 2023), we model the system-, user- and interface-dependencies of explanations of AI systems. ... We leverage the explanation type - explainer method mappings in the EO to identify what explainers to run."

    The central routing decision of the framework—which explainer method corresponds to which explanation type, and which template should structure the output—is loaded entirely from the Explanation Ontology, a resource created by the same authors in prior work. The paper does not independently validate these mappings against external ground truth, and its evaluations reuse the same EO-driven explanation types and templates. This makes the correctness of the user-centered explanation selection structurally dependent on the authors' own prior ontology.

full rationale

Most per-stage evaluations are self-contained: Decompose is scored against a human-verified question/reframed-question bank, Delegate uses standard explainability metrics (faithfulness, diversity, fidelity) on SHAP/DiCE/RuleXAI outputs, and Synthesis uses RAGAS metrics. These are not circular in the equation-level sense, and the user study provides external, if small-scale, evidence for subjective quality. However, the paper's own numbers create a concrete reduction problem for Contrastive explanations: Table 7 gives 0.00 F1 for Contrastive type identification, while Table 12 reports user satisfaction for explanations labeled 'Predicted Explanation Type: Contrastive.' Since a zero-precision/recall model cannot produce the required type label, the Contrastive evaluations must have used the gold type as an input, making the evaluation of that type an oracle-driven component test rather than an end-to-end prediction. In addition, the explanation-type-to-explainer and explanation-type-to-template mappings come from the authors' Explanation Ontology, cited as Chari et al. (May 2023) and never independently validated here; this is load-bearing self-citation because Delegate and Synthesis are defined by those mappings. The presence of independent component metrics and an external user study prevents a higher score, but the Contrastive issue is a genuine partial reduction, and the ontology dependence is a structural circularity risk.

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

The paper relies on the Explanation Ontology (created by the same authors) and on a chosen intermediate representation. No new physical or conceptual entities are introduced; the ontology is reused from prior work. The free-parameter list is empty because the central claim depends on the framework architecture and training choices, not on a small set of fitted constants.

assumptions (3)
  • domain assumption Explanation Ontology correctly represents user-centered explanation types and their mappings to explainer methods.
    Used in Delegate and Synthesis stages (Sections 2.2 and 3.1) to select explainers and templates; if the mapping is wrong, the wrong explainer runs.
  • ad hoc to paper Predicate logic is an adequate intermediate representation for user questions.
    Chosen in Decompose (Section 3.1) without comparison to other formats; the framework's ability to parse questions depends on this representation.
  • standard math Standard metrics (F1, faithfulness, context-utilization) measure explanation quality.
    Metrics are from cited literature (van der Waa et al. 2021, Es et al. 2023) and are applied without adaptation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MetaExplainer: A Framework to Generate Multi-Type User-Centered Explanations for AI Systems." pith.science (2026). https://pith.science/paper/JJGCGV6A

@misc{pith2026250800300,
  author       = {Pith},
  title        = {Pith review of: MetaExplainer: A Framework to Generate Multi-Type User-Centered Explanations for AI Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JJGCGV6A}},
  note         = {Machine review of arXiv:2508.00300}
}
read the original abstract

Explanations are crucial for building trustworthy AI systems, but a gap often exists between the explanations provided by models and those needed by users. To address this gap, we introduce MetaExplainer, a neuro-symbolic framework designed to generate user-centered explanations. Our approach employs a three-stage process: first, we decompose user questions into machine-readable formats using state-of-the-art large language models (LLM); second, we delegate the task of generating system recommendations to model explainer methods; and finally, we synthesize natural language explanations that summarize the explainer outputs. Throughout this process, we utilize an Explanation Ontology to guide the language models and explainer methods. By leveraging LLMs and a structured approach to explanation generation, MetaExplainer aims to enhance the interpretability and trustworthiness of AI systems across various applications, providing users with tailored, question-driven explanations that better meet their needs. Comprehensive evaluations of MetaExplainer demonstrate a step towards evaluating and utilizing current state-of-the-art explanation frameworks. Our results show high performance across all stages, with a 59.06% F1-score in question reframing, 70% faithfulness in model explanations, and 67% context-utilization in natural language synthesis. User studies corroborate these findings, highlighting the creativity and comprehensiveness of generated explanations. Tested on the Diabetes (PIMA Indian) tabular dataset, MetaExplainer supports diverse explanation types, including Contrastive, Counterfactual, Rationale, Case-Based, and Data explanations. The framework's versatility and traceability from using ontology to guide LLMs suggest broad applicability beyond the tested scenarios, positioning MetaExplainer as a promising tool for enhancing AI explainability across various domains.

Figures

Figures reproduced from arXiv: 2508.00300 by the authors.

Figure 1
Figure 1. Workflow diagram of MetaExplainer, highlighting the different input and output [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Instruction prompt example used for fine-tuning Llama models to decompose [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Decompose and Delegate outputs from the MetaExplainer for an example question [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Natural-language Explanations of Diabetes Prediction from Synthesis stage of [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Distribution of Likert ratings for overall system performance. (left) shows the dis [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 30 canonical work pages

  1. [1]

    Selecting predicate logic for knowledge representation by comparative study of knowledge representation schemes

    Amjad Ali and Mohammad Abid Khan. Selecting predicate logic for knowledge representation by comparative study of knowledge representation schemes. In 2009 international conference on emerging technologies, pages 23--28. IEEE, 2009

    work page 2009

  2. [2]

    One explanation does not fit all: A toolkit and taxonomy of ai explainability techniques, 2019

    Vijay Arya, Rachel KE Bellamy, Pin-Yu Chen, Amit Dhurandhar, Michael Hind, Samuel C Hoffman, Stephanie Houde, Q Vera Liao, Ronny Luss, Aleksandra Mojsilovi \'c , et al. One explanation does not fit all: A toolkit and taxonomy of ai explainability techniques, 2019. arXiv1909.03012

    arXiv 2019

  3. [3]

    Ai explainability 360: Impact and design

    Vijay Arya, Rachel KE Bellamy, Pin-Yu Chen, Amit Dhurandhar, Michael Hind, Samuel C Hoffman, Stephanie Houde, Q Vera Liao, Ronny Luss, Aleksandra Mojsilovi \'c , et al. Ai explainability 360: Impact and design. In Proc. of the Assoc. for the Adv. of Artif. Intell. (AAAI) Conf. on Artif. Intell., volume 36, pages 12651--12657, 2022

    work page 2022

  4. [4]

    Pima indians diabetes mellitus classification based on machine learning (ml) algorithms

    Victor Chang, Jozeene Bailey, Qianwen Ariel Xu, and Zhili Sun. Pima indians diabetes mellitus classification based on machine learning (ml) algorithms. Neural Compu. and Appl., 35 0 (22), Mar. 2023

    work page 2023

  5. [5]

    Informing clinical assessment by contextualizing post-hoc explanations of risk prediction models in type-2 diabetes

    Shruthi Chari, Prasant Acharya, Daniel M Gruen, Olivia Zhang, Elif K Eyigoz, Mohamed Ghalwash, Oshani Seneviratne, Fernando Suarez Saiz, Pablo Meyer, Prithwish Chakraborty, et al. Informing clinical assessment by contextualizing post-hoc explanations of risk prediction models in type-2 diabetes. Artif. Intell. in Med., 137: 0 102498, Feb. 2023

    work page 2023

  6. [6]

    Explanation ontology: A general-purpose, semantic representation for supporting user-centered explanations

    Shruthi Chari, Oshani Seneviratne, Mohamed Ghalwash, Sola Shirai, Daniel M Gruen, Pablo Meyer, Prithwish Chakraborty, and Deborah L McGuinness. Explanation ontology: A general-purpose, semantic representation for supporting user-centered explanations. Semantic Web J., Pre-press: 0 1--31, May 2023

    work page 2023

  7. [7]

    Boolean decision rules via column generation

    Sanjeeb Dash, Oktay Gunluk, and Dennis Wei. Boolean decision rules via column generation. In Advances in Neural Inf. Process. Syst. (NeurIPS), pages 4660--4670, 2018

    work page 2018

  8. [8]

    Human-centered explainability for life sciences, healthcare, and medical informatics

    Sanjoy Dey, Prithwish Chakraborty, Bum Chul Kwon, Amit Dhurandhar, Mohamed Ghalwash, Fernando J Suarez Saiz, Kenney Ng, Daby Sow, Kush R Varshney, and Pablo Meyer. Human-centered explainability for life sciences, healthcare, and medical informatics. Patterns, 3 0 (5), May 2022

    work page 2022

Show all 38 references
  1. [9]

    Accountability of ai under the law: The role of explanation, 2017

    Finale Doshi-Velez, Mason Kortz, Ryan Budish, Chris Bavitz, Sam Gershman, David O'Brien, Stuart Schieber, James Waldo, David Weinberger, and Alexandra Wood. Accountability of ai under the law: The role of explanation, 2017. arXiv1711.01134

    2017 arXiv

  2. [10]

    Ragas: Automated evaluation of retrieval augmented generation, 2023

    Shahul Es, Jithin James, Luis Espinosa-Anke, and Steven Schockaert. Ragas: Automated evaluation of retrieval augmented generation, 2023. arXiv2309.15217

    2023 arXiv

  3. [11]

    Retrieval-augmented generation for large language models: A survey, 2023

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey, 2023. arXiv2312.10997

    2023 arXiv

  4. [12]

    The false hope of current approaches to explainable artificial intelligence in health care

    Marzyeh Ghassemi, Luke Oakden-Rayner, and Andrew L Beam. The false hope of current approaches to explainable artificial intelligence in health care. The Lancet Digital Health, 3 0 (11): 0 e745--e750, Nov. 2021

    2021

  5. [13]

    Designing for ai explainability in clinical context

    Daniel M Gruen, Shruthi Chari, , Morgan A Foreman, Oshani Seneviratne, Rachel Richesson, Amar K Das, and Deborah L McGuinness. Designing for ai explainability in clinical context. In Trustworthy AI for Healthcare Workshop at Assoc. for the Adv. of Artif. Intell. (AAAI) 2021, p...

    2021

  6. [14]

    Supervised contrastive learning for pre-trained language model fine-tuning, 2020

    Beliz Gunel, Jingfei Du, Alexis Conneau, and Ves Stoyanov. Supervised contrastive learning for pre-trained language model fine-tuning, 2020. arXiv2011.01403

    2020 arXiv

  7. [15]

    Metrics for explainable ai: Challenges and prospects, 2018

    Robert R Hoffman, Shane T Mueller, Gary Klein, and Jordan Litman. Metrics for explainable ai: Challenges and prospects, 2018. arXiv1812.04608

    2018 arXiv

  8. [16]

    Lora: Low-rank adaptation of large language models, 2021

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. arXiv2106.09685

    2021 arXiv

  9. [17]

    Towards bridging the gaps between the right to explanation and the right to be forgotten

    Satyapriya Krishna, Jiaqi Ma, and Himabindu Lakkaraju. Towards bridging the gaps between the right to explanation and the right to be forgotten. In Int. Conf. on Mach. Learn. (ICML), pages 17808--17826, 2023

    2023

  10. [18]

    Rethinking explainability as a dialogue: A practitioner's perspective, 2022

    Himabindu Lakkaraju, Dylan Slack, Yuxin Chen, Chenhao Tan, and Sameer Singh. Rethinking explainability as a dialogue: A practitioner's perspective, 2022. arXiv2202.01875

    2022 arXiv

  11. [19]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Inf. Process. Syst....

    2020

  12. [20]

    Questioning the ai: informing design practices for explainable ai user experiences

    Q Vera Liao, Daniel Gruen, and Sarah Miller. Questioning the ai: informing design practices for explainable ai user experiences. In Proc. of the 2020 Conf. on Human Factors in Comp. Syst. (CHI), pages 1--15, 2020

    2020

  13. [21]

    Connecting algorithmic research and usage contexts: A perspective of contextualized evaluation for explainable ai

    Q Vera Liao, Yunfeng Zhang, Ronny Luss, Finale Doshi-Velez, and Amit Dhurandhar. Connecting algorithmic research and usage contexts: A perspective of contextualized evaluation for explainable ai. In Proc. of the Assoc. for the Adv. of Artif. Intell. (AAAI) Conf. on Human Compu...

    2022

  14. [22]

    Llamaindex

    LlamaIndex AI . Llamaindex. Accessed: Jun. 1, 2024 [Online]. Available: https://www.llamaindex.ai, 2023

    2024

  15. [23]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Infor. Process. Syst., pages 4768 -- 4777, 2017

    2017

  16. [24]

    Explaining answers from the semantic web: The inference web approach

    Deborah L McGuinness and Paulo Pinheiro Da Silva. Explaining answers from the semantic web: The inference web approach. J. of Web Semantics, 1 0 (4): 0 397--413, Oct. 2004

    2004

  17. [25]

    Explaining task processing in cognitive assistants that learn

    Deborah L McGuinness, Alyssa Glass, Michael Wolverton, and Paulo Pinheiro Da Silva. Explaining task processing in cognitive assistants that learn. In Assoc. for the Adv. of Artif. Intell. (AAAI) Spring Symp.: Interaction Challenges for Intell. Assistants, pages 80--87, 2007

    2007

  18. [26]

    Explanation in artificial intelligence: Insights from the social sciences

    Tim Miller. Explanation in artificial intelligence: Insights from the social sciences. Artif. Intell., 267: 0 1--38, Feb. 2019

    2019

  19. [27]

    Explaining explanations in ai

    Brent Mittelstadt, Chris Russell, and Sandra Wachter. Explaining explanations in ai. In Proc. of the Conf. on Fairness, Accountability, and Transparency, pages 279--288, 2019

    2019

  20. [28]

    Explaining machine learning classifiers through diverse counterfactual explanations

    Ramaravind K Mothilal, Amit Sharma, and Chenhao Tan. Explaining machine learning classifiers through diverse counterfactual explanations. In Proc. of the Conf. on Fairness, Accountability, and Transparency, pages 607--617, 2020

    2020

  21. [29]

    Why should i trust you?: Explaining the predictions of any classifier

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the predictions of any classifier. In Proc. of the 22nd ACM SIGKDD Int. Conf. on Knowl. Discovery and Data Mining, pages 1135--1144, 2016

    2016

  22. [30]

    Evaluating large language models in semantic parsing for conversational question answering over knowledge graphs

    Phillip Schneider, Manuel Klettner, Kristiina Jokinen, Elena Paslaru Bontas Simperl, and Florian Matthes. Evaluating large language models in semantic parsing for conversational question answering over knowledge graphs. In Int. Conf. on Agents and Artif. Intell. (IAAS), page 1...

    2024

  23. [31]

    Explaining machine learning models with interactive natural language conversations using talktomodel

    Dylan Slack, Satyapriya Krishna, Himabindu Lakkaraju, and Sameer Singh. Explaining machine learning models with interactive natural language conversations using talktomodel. Nature Mach. Intell., 5 0 (8): 0 873--883, Jul. 2023

    2023

  24. [32]

    Using the adap learning algorithm to forecast the onset of diabetes mellitus

    Jack W Smith, James E Everhart, WC Dickson, William C Knowler, and Robert Scott Johannes. Using the adap learning algorithm to forecast the onset of diabetes mellitus. In Proc. of the Annu. Symp. on Comp. Appl. in Med. Care, page 261. American Medical Informatics Association, 1988

    1988

  25. [33]

    Stanford alpaca: An instruction-following llama model

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model. Accessed Jul. 30, 2024 [Online]. Avaible: Stanford alpaca: An instruction-following llama model, 2023

    2024

  26. [34]

    What clinicians want: contextualizing explainable machine learning for clinical end use

    Sana Tonekaboni, Shalmali Joshi, Melissa D McCradden, and Anna Goldenberg. What clinicians want: contextualizing explainable machine learning for clinical end use. In Mach. Learn. for Health. Conf. (ML4HC), pages 359--380, 2019

    2019

  27. [35]

    Evaluating xai: A comparison of rule-based and example-based explanations

    Jasper van der Waa, Elisabeth Nieuwburg, Anita Cremers, and Mark Neerincx. Evaluating xai: A comparison of rule-based and example-based explanations. Artif. Intell., 291, 2021

    2021

  28. [36]

    Counterfactual explanations without opening the black box: Automated decisions and the gdpr

    Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the gdpr. Harv. JL & Tech., 31, Mar. 2018

    2018

  29. [37]

    Designing theory-driven user-centric explainable ai

    Danding Wang, Qian Yang, Ashraf Abdul, and Brian Y Lim. Designing theory-driven user-centric explainable ai. In Proc. of the 2019 Conf. on Human Factors in Comp. Syst. (CHI), pages 1--15, 2019

    2019

  30. [38]

    Evaluating the quality of machine learning explanations: A survey on methods and metrics

    Jianlong Zhou, Amir H Gandomi, Fang Chen, and Andreas Holzinger. Evaluating the quality of machine learning explanations: A survey on methods and metrics. Electronics, 10 0 (5), Mar. 2021

    2021

Pith tools

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