Pith. sign in

REVIEW 3 major objections 5 minor 18 references

EMMM, Explain Me My Model! Explainable Machine Generated Text Detection in Dialogues

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read EMMM is an explanation-then-detection framework that turns raw attribution scores into natural-language reports and dialogue-act-conditioned visuals for conversational machine-generated text, preferred by 69–70% of human evaluators while ke

desk verdict Useful framework for explainable MGT detection in dialogues, with a genuinely interesting DA-based semi-global aggregation and open code/data, but the sub-second latency claim is contradicted by the paper's own timing table. read the letter →

arxiv 2508.18715 v1 pith:AV5B3SBE submitted 2025-08-26 cs.CL

Angela Yifei Yuan , Haoyi Li , Soyeon Caren Han , Christopher Leckie This is my paper · ORCID
classification cs.CL
keywords machine-generatedtextdetectionexplainableAIdialogueactsconversationalnaturallanguageexplanationssemi-globalaggregationreal-time
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 machine-generated text detection in live customer-service dialogues can be made accurate, fast, and explainable to non-experts at the same time. The proposed EMMM framework first extracts dialogue acts—the intent and slot structure of each user utterance—then runs turn-level and dialogue-level detection, and finally explains each decision through a template-based natural-language report plus a semi-global visualization that aggregates attributions by dialogue act. On two dialogue benchmarks, EMMM's Macro-F1 averages 0.9858, above most fully supervised baselines; in a human survey, 69% of responses preferred EMMM's explanation report over a raw attribution listing, with an even stronger preference (71%) among participants without an AI background. The authors claim this is the first framework to make conversational MGT detection interpretable for non-expert users while remaining suitable for near-real-time deployment.

What carries the argument

Dialogue acts—the intent-and-slot labels of each utterance, such as inform-hotel-area—serve as both detection features and the pivot for explanation. The mechanism carrying the argument is dialogue-act-conditioned semi-global aggregation: attributions are accumulated per dialogue act per class over a training set, and at inference the target utterance's acts retrieve and sum the relevant importance profiles, yielding a global-but-contextualized picture that the authors show is more relevant to local predictions than plain global aggregation. A lightweight template renders the top acts and tokens as natural-language sentences, and a sequential selector–predictor pipeline limits explanation co

What would settle it

Measure end-to-end latency on a new domain with a supervised dialogue-act extractor: if mean per-utterance time exceeds one second—or if shuffling the extracted dialogue acts fails to reduce the AOPC advantage over global aggregation—the paper's real-time and relevance claims are falsified.

Watch

Extended reading notes

Core claim

EMMM is, to the authors' knowledge, the first framework for explainable machine-generated text detection in conversational settings aimed at non-experts. Its central claim is that dialogue acts, grounded in speech act theory, give explanations a behavioral dimension that raw token attributions lack. The framework detects at two levels—per turn and across the whole dialogue—using a sequential selector that keeps only the most influential tokens and acts, and it explains with a lightweight natural-language template and a dialogue-act-contextualized semi-global visualization. The authors report that this combination yields state-of-the-art detection accuracy (average Macro-F1 0.9858 across SPAD

Load-bearing premise

The 'under one second' latency claim assumes dialogue-act extraction is fast and accurate on the target domain, but the few-shot LLM extractor used for the Frames dataset takes about 3.5 seconds per utterance.

Editorial extensions

If this is right

  • Detection stays accurate under strict interpretability budgets: using only three dialogue acts and three tokens per utterance, EMMM beats zero-shot and pre-trained supervised baselines on both datasets.
  • Dialogue-act-conditioned aggregation yields explanation relevance scores (AOPC) that are significantly higher than global aggregation for both AI and human classes (p < 0.05).
  • Human evaluators, especially non-experts, prefer EMMM's explanation reports to raw attribution listings; the survey reports 69% overall and 71% among participants without AI background.
  • With a fast supervised dialogue-act extractor, report generation completes in under half a second and the whole explanation pipeline in about one second per utterance, making online deployment plausible.
  • The framework is modular: swapping attribution methods or base models keeps Macro-F1 above 0.9, suggesting the explanation scheme is not tied to one detector.

Reading between the lines

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

  • If dialogue-act extraction can be made fast and accurate on new domains, the same selector–predictor design should transfer to other asymmetric detection tasks, such as catching scripted or synthetic agent messages.
  • The latency claim is only as strong as the DA extractor: with few-shot LLM extraction on Frames taking 3.5 seconds per utterance, the 'under one second' promise should be read as applying to the explanation stages, not the full pipeline, unless DA extraction is replaced by a supervised model.
  • DA-conditioned aggregation is a testable hypothesis about human judgment: corrupting or shuffling dialogue acts should reduce both AOPC relevance and user preference, confirming that DAs, not the template, drive the explanation's value.
  • The 13-participant survey is a preference signal rather than a decision-quality measurement; a field study asking moderators to find real machine-generated turns with and without EMMM reports would tell whether the preference translates into faster, more accurate review.
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 proposes EMMM, an explainable machine-generated text detection framework for dialogues. It combines turn-level and dialogue-level detection, uses dialogue acts as behavioral features, generates natural-language and semi-global visual explanations, and evaluates on SPADE plus a new synthetic Frames dataset. The authors report competitive Macro-F1 scores, higher AOPC relevance for DA-conditioned explanations, a 13-participant preference survey, and per-step runtime measurements.

Significance. If the claims hold, EMMM is a useful step toward deployable, non-expert-interpretable MGT detection in customer-service dialogues. Strengths include the explicit integration of dialogue acts into detection and explanation, the semi-global aggregation idea, the reproducible open-source release, and the use of multiple attribution baselines. The AOPC comparison is backed by a paired significance test. However, the headline efficiency and interpretability claims currently exceed the evidence, and the accuracy claim is overstated relative to the reported MLP baseline.

major comments (3)
  1. [Abstract, Section 1, Section 5 (Table 3)] The real-time claim is internally inconsistent. The abstract says 'generating outputs within 1 second' and Section 1 says EMMM 'produces explanation reports online in under 1 second,' but Table 3 reports full-framework times of 2.19 s for SPADE and 4.70 s for Frames. The bottleneck is DA extraction, which takes 3.48 s per utterance for Frames under the Qwen2.5-7B few-shot setup. The suggestion that this step 'can be replaced by a supervised model' is not demonstrated for Frames, so the efficiency claim is not established for the dataset introduced to show domain coverage. Please either benchmark a fast DA extractor on Frames or revise the latency claim to refer only to the explanation-report subprocess (steps 4 and 7), and adjust the abstract accordingly.
  2. [Section 5, Interpretability; Conclusion] The interpretability claim rests on a small survey: 13 participants, 52 responses, four samples, with no significance testing or inter-rater variability analysis. The conclusion states that EMMM's explanations 'significantly improve user comprehension,' but this is not supported by any statistical test. Add a significance test (e.g., exact binomial test on per-response preference) and discuss the small sample, or temper the claim to a participant preference trend. Also note the abstract says 70% while the body and Figure 5 report 69%; please reconcile.
  3. [Section 5, Table 1 and following text] The accuracy claim 'state-of-the-art' is overstated. In Table 1, the fully supervised MLP baseline achieves Macro-F1 0.9906 on SPADE and 0.9976 on Frames, both higher than EMMM's 0.9771 and 0.9945. EMMM is competitive with the best supervised methods, but it is not the top performer on either dataset. Please rephrase the contribution and abstract to say 'competitive accuracy' rather than state-of-the-art, and discuss the small margin relative to the MLP baseline, which is relevant to the cost of interpretability.
minor comments (5)
  1. [Abstract vs. Section 5] The preference percentage is 70% in the abstract and 69% in Section 5 and Figure 5. Please make these consistent.
  2. [References] The reference 'OW ASP 2025' appears with an odd space; fix to 'OWASP' and ensure the URL is correct.
  3. [Abstract] Minor grammar: 'explanation become crucial' should be 'explanations become crucial'.
  4. [Table 7] The hyperparameter row 'Number of tree' should be 'Number of trees'.
  5. [Section 5, Time Complexity] The phrase 'under one second' is used for the 'overall explanation process' but readers may confuse it with end-to-end framework time. Please define clearly whether the sub-second claim refers to steps 4+7 only or to the complete pipeline including DA extraction.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: EMMM's detection/explanation pipeline is evaluated against independent baselines and standard faithfulness metrics; the SPADE self-citation is a dataset contribution, not a load-bearing derivational step.

full rationale

The paper does not derive its central claims from its own premises by construction. The detection model (DistilGPT2) is trained on the SPADE and Frames dialogue data and compared against external zero-shot, pretrained-supervised, and fully supervised baselines (Binoculars, RADAR, ChatGPTDroberta, MLP, etc.); no parameter fitted to the evaluation target is subsequently relabeled as a prediction. The explanation pipeline uses standard attribution methods (Faith-SHAP, STII, Integrated Gradients) and evaluates faithfulness via AOPC; the fact that aggregated features are derived from the same model is the standard explanatory-faithfulness test, not a tautology. The only notable self-citation is SPADE (Li et al. 2025), which is used as a benchmark and as a generation framework for the Frames extension; this does not constitute a circular step because the dataset is an external artifact of prior work rather than an assumption equivalent to the present conclusions. No uniqueness theorem is imported from the authors to force a choice, and no ansatz is smuggled in via citation: the sequential selector–predictor design cites Luo et al. (2024), an independent survey. The paper's abstract claim of 'outputs within 1 second' is internally inconsistent with Table 3 (framework totals of 2.1921s for SPADE and 4.6966s for Frames), but an internal inconsistency is a correctness risk, not a circularity. Accordingly, no circular step is identified; score 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework does not introduce new physical or mathematical entities; the only introduced artifacts are a dataset and a template-based explanation system. The free parameters listed are hand-chosen or validation-tuned features central to the selector-predictor design.

free parameters (3)
  • feature budget = 3 dialogue acts + 3 tokens per utterance
    Selected on validation set across attribution methods; impacts detection/explanation trade-off (Table 2).
  • CSLS match threshold coefficient = 0.9
    Used in text-DA matching (Appendix A, step 7); chosen to balance matches.
  • phrase merge overlap K = 2
    Used to de-duplicate phrase display in word clouds (Appendix A).
assumptions (4)
  • domain assumption Speech act theory (Austin 1975) is a valid basis for using dialogue acts as behavioral features.
    The framework's multi-dimensional design and semi-global aggregation depend on dialogue acts capturing meaningful intent; no empirical validation is provided in this paper.
  • domain assumption Faith-SHAP attribution scores faithfully reflect the turn-level model's feature importance.
    The selector picks the highest absolute attribution features; if attributions are unreliable, the dialogue-level predictor sees noisy features.
  • domain assumption The synthetic Frames dataset generated with Qwen2.5-32B is representative of real AI chatbot dialogues.
    The second benchmark is generated with the same LLM-generation framework; results may not transfer to other chatbot behaviors.
  • domain assumption Human preference for an explanation implies better comprehension.
    The survey asks for preference, not whether users could perform a task correctly with the explanation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EMMM, Explain Me My Model! Explainable Machine Generated Text Detection in Dialogues." pith.science (2026). https://pith.science/paper/AV5B3SBE

@misc{pith2026250818715,
  author       = {Pith},
  title        = {Pith review of: EMMM, Explain Me My Model! Explainable Machine Generated Text Detection in Dialogues},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AV5B3SBE}},
  note         = {Machine review of arXiv:2508.18715}
}
read the original abstract

The rapid adoption of large language models (LLMs) in customer service introduces new risks, as malicious actors can exploit them to conduct large-scale user impersonation through machine-generated text (MGT). Current MGT detection methods often struggle in online conversational settings, reducing the reliability and interpretability essential for trustworthy AI deployment. In customer service scenarios where operators are typically non-expert users, explanation become crucial for trustworthy MGT detection. In this paper, we propose EMMM, an explanation-then-detection framework that balances latency, accuracy, and non-expert-oriented interpretability. Experimental results demonstrate that EMMM provides explanations accessible to non-expert users, with 70\% of human evaluators preferring its outputs, while achieving competitive accuracy compared to state-of-the-art models and maintaining low latency, generating outputs within 1 second. Our code and dataset are open-sourced at https://github.com/AngieYYF/EMMM-explainable-chatbot-detection.

Figures

Figures reproduced from arXiv: 2508.18715 by the authors.

Figure 1
Figure 1. A demonstration of EMMM framework online detection and non-expert oriented explanation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The 7 steps process of the EMMM explainable detection framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of explanation relevance between semi-global ( [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: An example comparison of contextualized semi [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Human survey result on user interpretability pref [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison of Macro-F1 scores using different [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: A demonstration of baseline attribution-based ex [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 15 canonical work pages

  1. [1]

    {intent or intent description } {slot and optionally (slot description)} {value}

    Convert each DAi (intent, domain, slot, value) to a text string TDA,i in format “ {intent or intent description } {slot and optionally (slot description)} {value}”. For ex- ample, (inform, travel, or city, Gotham City)is converted to “inform or city (Origin city) Gotham City ”. Whereas non-natural language intents like “nobook” is replaced by their descri...

    work page

  2. [2]

    Encoder inference on entire utterance ( Tutt), and on each individual DA text ( TDA,i ∈ DAutt). This study uses a paraphrase-MiniLM-L6-v2 model (Reimers and Gurevych 2019) fine-tuned for 3 epochs on a balanced dataset of positive and negative samples using CosineS- imilarityLoss. For each utterance in the training set, pos- itive samples have it paired wi...

    work page 2019

  3. [3]

    Extract token embeddings for each utterance token (tutt ∈ Tutt), and for each DA token (tDA ∈ TDA,i)

    work page

  4. [4]

    Calculate cosine similarity for each pair of utterance and DA tokens: {cos sim(tutt, tDA)|tutt ∈ Tutt, tDA ∈ TDA,i, TDA,i ∈ DAutt}

    work page

  5. [5]

    Artificial Intelligence Review , 57: 265

    Contrasting Linguistic Patterns in Human and LLM- Generated News Text. Artificial Intelligence Review , 57: 265. OW ASP. 2025. OW ASP Top 10 for Large Language Model Applications. https://owasp.org/www-project-top-10-for- large-language-model-applications/. Accessed: 2025-04- 06. Qwen Team. 2024. Qwen2.5: A Party of Foundation Mod- els. https://qwenlm.git...

    arXiv 2025

  6. [6]

    2018) adjusts similarity scores by increas- ing those for features with few close neighbors and de- creasing those for features with many

    Cross-Domain Similarity Local Scaling (CSLS) (Lam- ple et al. 2018) adjusts similarity scores by increas- ing those for features with few close neighbors and de- creasing those for features with many. This reduces the influence of tokens that are broadly similar to many DAs, or vice versa, ensuring more relevant matches are prioritized. The CSLS-adjusted ...

    work page 2018

  7. [7]

    looking forward to it!

    For each utterance token tutt, match it with any DA TDA,i that satisfies the condition: scsls(tutt, TDA,i) ≥ mintutt + 0.9 × (maxtutt − mintutt ), where mintutt and maxtutt denote the minimum and maximum scsls the token tutt has across all TDA,i ∈ DAutt. If a DA is matched to continuous text spans, n-gram phrases can be extracted. To enable investigation ...

    work page

  8. [10]

    Similarity between an utterance token and an entire DA is defined as the maximum similarity score be- tween the utterance token with each DA token within the DA: s(tutt, TDA,i) = max{cos sim(tutt, tDA)|tDA ∈ TDA,i}

    work page

Show all 18 references
  1. [13]

    looking forward to it!

    Filter duplicates : retain phrases that either (1) do not have any other phrases containing it or (2) do not con- tain any other phrases. For example, “looking forward to it!” and “forward” would be returned, whereas “looking forward” would be removed to reduce duplications

  2. [14]

    I’m looking forward to

    Merge phrase: merge any two phrases if there exists a consecutive overlap from the 2 ends for a minimum of K = 2 tokens. For example, “I’m looking forward to” would be merged with “looking forward to it!” as “I’m looking forward to it!” to further reduce replications. B Experi...

    2025

  3. [15]

    The initial request from the user, or

  4. [16]

    If nothing matches your constraints,

    An alternative suggested by the user after the system fails to meet the previous goal. If a goal was unsuccessful, the user either ended the dialogue or continued with an **alternative goal**, which must begin with: “If nothing matches your constraints, ...” Please differentia...

  5. [17]

    Options within the same goal: The user modifies previously specified constraints voluntarily to explore and compare different options, even when the system has already returned packages that match their earlier constraints

  6. [18]

    If nothing matches your constraints,

    Alternative goal: The user modifies constraints as a fallback because the system was unable to find any matching packages with the original constraints. This must start with “If nothing matches your constraints, ...”. Alternative goals can also include multiple options within ...

  7. [640]

    Eric, M.; Goel, R.; Paul, S.; Kumar, A.; Sethi, A.; Goyal, A

    Springer. Eric, M.; Goel, R.; Paul, S.; Kumar, A.; Sethi, A.; Goyal, A. K.; Ku, P.; Agarwal, S.; Gao, S.; and Hakkani-T ¨ur, D

  8. [2020]

    In 12th International Conference on Language Re- sources and Evaluation, LREC 2020 , 422–428

    MultiWOZ 2.1: A consolidated multi-domain dia- logue dataset with state corrections and state tracking base- lines. In 12th International Conference on Language Re- sources and Evaluation, LREC 2020 , 422–428. European Language Resources Association (ELRA). Gehrmann, S.; Strob...

    2020 arXiv

  9. [2022]

    In Carpuat, M.; de Marneffe, M.-C.; and Meza Ruiz, I

    Few-Shot Self-Rationalization with Natural Lan- guage Prompts. In Carpuat, M.; de Marneffe, M.-C.; and Meza Ruiz, I. V ., eds., Findings of the Association for Computational Linguistics: NAACL 2022, 410–424. Seattle, United States: Association for Computational Linguistics. Mo...

    2022

  10. [2024]

    AI vs Humans

    Spotting LLMs with binoculars: zero-shot detec- tion of machine-generated text. In Proceedings of the 41st International Conference on Machine Learning , ICML’24. JMLR.org. Hu, X.; Chen, P.; and Ho, T. 2023. RADAR: Robust AI-Text Detection via Adversarial Learning. In Advances...

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.