Joint Effects of Argumentation Theory, Audio Modality and Data Enrichment on LLM-Based Fallacy Classification

Hongxu Zhou; Hylke Westerdijk; Khondoker Ittehadul Islam

REVIEW 4 major objections 4 minor 1 cited by

Joint Effects of Argumentation Theory, Audio Modality and Data Enrichment on LLM-Based Fallacy Classification

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Adding context and audio emotion metadata usually lowers LLM fallacy-classification accuracy, biasing labels toward Appeal to Emotion.

desk verdict Useful exploratory study with a clear negative finding about audio metadata, but the headline claims outrun the 120-example single-run evidence. read the letter →

arxiv 2509.11127 v1 pith:2TCLJO2P submitted 2025-09-14 cs.CL

Hongxu Zhou , Hylke Westerdijk , Khondoker Ittehadul Islam This is my paper · ORCID

classification cs.CL

keywords fallacyclassificationlargelanguagemodelschain-of-thoughtpromptingpragma-dialecticsperiodictableofargumentsmultimodalaudiometadatapresidentialdebatesattentiondilution

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 tries to establish that enriching LLM prompts with extra information—debate date context and audio-derived emotional tone—does not reliably help fallacy classification and often hurts it. Using six fallacy types from U.S. presidential debate snippets, the authors compare a basic prompt with two theory-driven Chain-of-Thought prompts and three input settings. They find that basic prompts often match or beat enriched ones, that context helps only for specific fallacy types, and that emotional tone metadata pushes the model to over-predict Appeal to Emotion at the expense of logical reasoning. The stakes are practical: automatic fallacy detection is a tool against manipulation and misinformation, but this result warns that naive data enrichment can backfire.

What carries the argument

The experimental design itself is the machinery: three prompt conditions (a basic instruction-and-definitions prompt, a Pragma-Dialectics prompt that asks the model to identify violated rules of critical discussion, and a Periodic Table of Arguments prompt that decomposes arguments into form, substance, and lever) are crossed with three input conditions (text-only, text plus debate date, text plus date plus audio-derived emotional tone values binned as low/moderate/high). The cross-comparison isolates what each added cue contributes, and the confusion matrices and qualitative traces expose how emotional metadata redirects the model's final label.

What would settle it

Run the same three prompts and three input conditions on a larger class-balanced test set with multiple random seeds and report confidence intervals; the central claim falls if context+audio does not reliably lower accuracy or shift the confusion matrix toward Appeal to Emotion.

Watch

Extended reading notes

Core claim

The central claim is that the extra inputs commonly assumed to aid LLM reasoning—surrounding debate context and the speaker's emotional tone—act as distractors rather than aids in fallacy classification. Across a class-balanced test set, the model's best overall accuracy came from the basic Pragma-Dialectics prompt without enrichment; adding context rarely helped broadly, and adding arousal/dominance/valence metadata shifted predictions toward Appeal to Emotion, degrading performance most severely under the Periodic Table of Arguments prompt. The paper proposes "attention dilution" as the mechanism: the model overweights salient emotional cues and loses sight of the statement's logical struc

Load-bearing premise

The load-bearing premise is that the observed accuracy differences between input conditions are caused by the input modifications, not by sampling noise or decoder randomness—yet the paper reports single-run point estimates on a 120-instance test set.

Editorial extensions

If this is right

If the paper is right, one-size-fits-all context enrichment is suboptimal for fallacy classification; adding metadata needs to be justified per fallacy type and prompt.
Audio emotional tone metadata should be treated as a bias source in multimodal fallacy detection, not just as extra signal, because it systematically inflates Appeal to Emotion predictions.
Theory-informed chain-of-thought prompts can improve interpretability and occasionally accuracy, but not enough to offset the harm of noisy enrichment.
Context can help for context-dependent fallacies such as Ad Hominem and False Cause, but the overall tradeoff is negative.
Slippery Slope is the hardest category across all conditions, and audio metadata can make it collapse entirely under structural reasoning.

Reading between the lines

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

Editorial inference: the "attention dilution" explanation could be tested directly by adding random or irrelevant metadata instead of emotional tone; if accuracy drops similarly, dilution rather than semantic emotion content is the active ingredient.
Editorial inference: because the results are point estimates from a single run on 120 examples, the reported differences between conditions may not be reliable; a replication with multiple seeds and confidence intervals is the natural next test.
Editorial inference: the binning of continuous arousal/valence/dominance values into low/moderate/high may discard useful information; feeding raw values or letting the model process the audio directly could alter or reverse the findings.

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

Colleague, quick read of arXiv:2509.11127. The takeaway: this is a cleanly designed exploratory study, and the central caution—that appending emotional-tone metadata to prompts biases Qwen-3-8B toward 'Appeal to Emotion' and hurts overall fallacy classification—is plausible but not actually established by the numbers. The paper is worth a serious referee, but only if the authors agree to reframe it as preliminary and add uncertainty quantification.

What's genuinely new: the combination of two theory-driven CoT prompt families (Pragma-Dialectics, Periodic Table of Arguments) with wav2vec2-derived arousal/dominance/valence metadata in a single-task LLM evaluation. I haven't seen that mix. The paper also does a few things well: it reports data and code, gives a clear setup, includes qualitative traces showing the same sentence flipping predictions across conditions, and it is honest in Section 6.1 that conclusions are preliminary. The 'attention dilution' mechanism is a reasonable hypothesis, not a proof, and they mostly keep it hedged there.

The soft spots are the load-bearing ones. The test set is 120 instances, 20 per class. No confidence intervals, no significance tests, no multiple seeds. Temperature is 0.6 with top-p/top-k set, so run-to-run variance is non-trivial. In Table 1, the base vs context+audio accuracy gap is 51% to 46%—six items out of 120. The per-class recall jumps that motivate the 'Appeal to Emotion bias' claim (e.g., 80→95 in Table 4, 65→90 in Table 8) involve three to five items per cell. Those are within binomial noise. The one large effect—PTA context+audio collapsing to 25% accuracy—looks real and important, but it's a single condition and could reflect an instruction-following collapse rather than a general 'attention dilution' mechanism. The abstract overstates what the data can support; the authors' own Section 6.1 is more careful.

I don't see circularity or an invented audit trail. The prompts don't encode test labels, and the free parameters (emotion thresholds, decoding settings) are few and honestly disclosed.

Who is this for: people building multimodal fallacy detectors, especially with generative models; shared-task participants; and anyone studying prompt modality interaction. It's a useful cautionary data point, not a definitive result.

Recommendation for peer review: send it out, but with an expectation of major revision—require confidence intervals or bootstrap, ideally multiple seeds, a sensitivity analysis on the emotion threshold, and a discussion that limits the claims to the single model and balanced subsample. Honestly, it would sit comfortably as a workshop paper or an extended abstract; as a journal article, the statistical foundations need work.

Referee Report

4 major / 4 minor

Summary. The paper studies LLM-based fallacy classification on a class-balanced subset of the MM-ArgFallacy2025 shared task, using Qwen-3-8B. It compares three prompt strategies (a basic prompt and two theory-grounded Chain-of-Thought prompts based on Pragma-Dialectics and the Periodic Table of Arguments) under three input conditions (text-only, text plus date context, and text plus context and wav2vec2-derived emotional tone metadata). The central claims are that theoretical prompting improves interpretability but not consistently accuracy, that adding context and especially emotional tone metadata often lowers performance, that emotional tone biases predictions toward Appeal to Emotion, and that basic prompts often outperform enriched ones. Data and code are made publicly available.

Significance. If the claims were statistically established, the paper would make a useful contribution to multimodal fallacy detection and to the literature on context and audio metadata in LLM prompting. The experimental design is clear and transparent: class-balanced validation and test sets, per-class metrics, confusion matrices, and qualitative reasoning traces. The authors also provide open code and data, which is a concrete strength. However, the evidence is currently preliminary: the test set contains only 120 instances (20 per class), and all results come from a single stochastic run with no confidence intervals or significance tests. The paper is therefore better read as a well-documented exploratory study than as a definitive demonstration of the reported effects.

major comments (4)

[§5, Tables 1, 5, 9 and §6.1] The central claims are supported only by point estimates from a single run on 120 test instances (20 per class). Differences such as base accuracy 51% vs 49% vs 46% (Table 1) and PD accuracy 53% vs 49% vs 46% (Table 5) correspond to 5–7 instances out of 120; the binomial standard error at n=120 is about 4.5 percentage points. Per-class recall shifts such as 80%→95% (Table 4) and 65%→90% (Table 8) correspond to 3–5 instances out of 20, with a per-class SE around 10 points. No confidence intervals, significance tests, or multi-seed averaging are reported. The authors' own §6.1 labels the conclusions 'preliminary,' yet the abstract presents them as general findings. The one large effect that is unlikely to be noise is the PTA context+audio collapse to 25% accuracy (Table 9), but that is a single condition and may reflect a prompt-instruction breakdown rather than the general 'attention dilu
[§4.4] The experiments use a single model (Qwen-3-8B) with temperature 0.6, top_p 0.95, and top_k 20, but no random seeds are reported and no repeated generations are performed. LLM decoding is stochastic at this temperature, so run-to-run variation can easily produce the observed 1–8 point accuracy differences. This undermines the load-bearing inference that differences across conditions are caused by the input modifications. At minimum, the authors should report results averaged over several seeds, provide confidence intervals, or explicitly restrict all conclusions to within-run qualitative comparisons.
[§6.1] The 'attention dilution' explanation is not tested. The paper does not measure attention, nor does it compare against controlled alternatives such as adding irrelevant text of the same length, adding the emotional tone values without categorical labels, or adding a different categorical cue. The observed degradation is consistent with other explanations, including the model following the provided low/moderate/high emotion categories too literally, or the PTA prompt becoming overly complex when extra metadata is appended. The mechanism should be reframed as a hypothesis, not a demonstrated conclusion.
[Abstract vs §6.1] The abstract states as a general result that 'emotional tone metadata biases the model toward labeling statements as Appeal to Emotion, worsening logical reasoning,' while §6.1 explicitly says 'This study's conclusions are preliminary.' This inconsistency is important because the evidence for the bias is mainly per-class recall increases over 20-instance cells. The abstract and conclusions should be reworded to reflect the statistical uncertainty and the exploratory nature of the study.

minor comments (4)

[Throughout] There are several typos and inconsistencies: 'challanges' (§3), 'condistions' (Table 10 caption), 'market dataset' (§4.2), inconsistent capitalization of 'MAMKit'/'mamkit', and a duplicated sentence about setting temperature to 0.6 (§4.4). Please proofread carefully.
[Appendix Figures] The confusion matrices and difference matrices (Figures 4–18) are referenced but not shown in the main text and the appendix appears to contain only captions. Please ensure the actual figures are included and readable, and refer to them in the body where the reader is expected to inspect them.
[§4.2] The 'context' condition is only the date of the debate. The claim that the model can infer the speaker and topic from the date is speculative. Scope the 'context' claims to 'date metadata' throughout, since this is a much weaker form of context than dialogue context.
[§5, per-class tables] Per-class percentages over 20 instances should be accompanied by raw counts, especially for cells showing 0% or 100%. For example, a precision of 100% for Ad Hominem in the PTA base condition (Table 11) may be based on one or two positive predictions; raw counts would prevent over-interpretation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the paper is an empirical comparison with externally sourced data and models, and its 'attention dilution' interpretation is an acknowledged post-hoc explanation rather than a fitted input.

full rationale

The paper's core claims—that context/audio metadata changes LLM fallacy classification and biases predictions toward Appeal to Emotion—are empirical observations from test-set evaluations, not consequences of definitions or fitted parameters. The prompts (Basic, PD, PTA) are constructed from argumentation theory sources (Van Eemeren & Grootendorst 1987; Wagemans 2023) and do not encode the test labels or the outcome measure. Audio features are extracted with a pre-trained wav2vec2 model (Wagner et al. 2023) and fed to the LLM as low/moderate/high categories; the mapping from feature values to categories is a fixed heuristic, not fitted to the test set. No parameter is estimated from the 120-instance test set, so the accuracy/recall differences are measured, not forced. There is no self-citation chain, uniqueness theorem, or ansatz imported from the authors' prior work; the references are to external prior work. The 'attention dilution' account is explicitly labeled preliminary in Section 6.1 and is an interpretation of observed results, not a premise used to generate them. Concerns about statistical reliability (single run, no confidence intervals) are validity/generalizability risks, not circularity. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central empirical comparison has no derivation, so the ledger captures the hand-chosen thresholds and decoding parameters, plus the domain assumptions about label quality, audio feature validity, and theory-to-label mapping.

free parameters (2)

Emotion category thresholds = ±0.33
Continuous arousal/dominance/valence values are mapped to low/moderate/high with a heuristic threshold of ±0.33 (Section 4.5.4), chosen by the authors.
Decoding parameters (temperature, top_p, top_k) = 0.6 / 0.95 / 20
Model generation settings fixed without systematic tuning (Section 4.4).

assumptions (4)

domain assumption The gold labels in the MAMKit dataset for the MM-ArgFallacy2025 classification task are correct.
Ground truth is taken as given from the shared task dataset (Section 4.1) without annotation quality checks.
domain assumption The wav2vec 2.0-large model of Wagner et al. (2023) yields valid arousal, dominance, and valence scores for these debate audio snippets.
Audio emotion features are extracted from a pre-trained model (Section 4.3) without validation on this dataset.
ad hoc to paper The Pragma-Dialectics and Periodic Table of Arguments frameworks can be faithfully mapped to the six fallacies used in the shared task.
The authors construct a mapping from theory rules to six specific fallacy types (Sections 4.5.2, 4.5.3), which can introduce distortion.
domain assumption The debate date alone is sufficient context for the model to infer speaker and topic information relevant to fallacy classification.
Section 4.2: 'we expect the model to infer details about the speaker and the topic of each debate' from the date metadata.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Joint Effects of Argumentation Theory, Audio Modality and Data Enrichment on LLM-Based Fallacy Classification." pith.science (2026). https://pith.science/paper/2TCLJO2P

@misc{pith2026250911127,
  author       = {Pith},
  title        = {Pith review of: Joint Effects of Argumentation Theory, Audio Modality and Data Enrichment on LLM-Based Fallacy Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2TCLJO2P}},
  note         = {Machine review of arXiv:2509.11127}
}

read the original abstract

This study investigates how context and emotional tone metadata influence large language model (LLM) reasoning and performance in fallacy classification tasks, particularly within political debate settings. Using data from U.S. presidential debates, we classify six fallacy types through various prompting strategies applied to the Qwen-3 (8B) model. We introduce two theoretically grounded Chain-of-Thought frameworks: Pragma-Dialectics and the Periodic Table of Arguments, and evaluate their effectiveness against a baseline prompt under three input settings: text-only, text with context, and text with both context and audio-based emotional tone metadata. Results suggest that while theoretical prompting can improve interpretability and, in some cases, accuracy, the addition of context and especially emotional tone metadata often leads to lowered performance. Emotional tone metadata biases the model toward labeling statements as \textit{Appeal to Emotion}, worsening logical reasoning. Overall, basic prompts often outperformed enhanced ones, suggesting that attention dilution from added inputs may worsen rather than improve fallacy classification in LLMs.

Figures

Figures reproduced from arXiv: 2509.11127 by the authors.

**Figure 5.** Confusion matrix for the Context condition [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 3.** Mean Audio Feature Distribution of Arousal, [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Confusion matrix for the Base condition un [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 8.** Figure 8: Difference matrix for the Context + Audio condition − Base condition under the Basic prompt. Pragma-Dialectic Confusion & Difference Matrices [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

**Figure 12.** Figure 12: Confusion matrix for the Context condition − Base condition under the PD prompt [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

**Figure 10.** Figure 10: Confusion matrix for the Context condition [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

**Figure 17.** Figure 17: Difference matrix for the Context condition [PITH_FULL_IMAGE:figures/full_fig_p013_17.png]

**Figure 18.** Figure 18: Difference matrix for the Context + Audio [PITH_FULL_IMAGE:figures/full_fig_p013_18.png]

**Figure 16.** Figure 16: Confusion matrix for the Context + Audio [PITH_FULL_IMAGE:figures/full_fig_p013_16.png]

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

Evaluating VLMs on Multimodal Aristotelian Persuasion Tasks
cs.CL 2026-08 conditional novelty 3.0 of 10

Qwen2-VL and Qwen3-VL beat 2022 ImageArg baselines in zero-shot Logos and Ethos F1, but Qwen3 falls below baseline on Pathos despite the abstract's claim.

Reference graph

Works this paper leans on

16 extracted references · cited by 1 Pith paper

[1]

Haddadan, V

Goffredo, P., S. Haddadan, V. Vorakitphan, E. Cabrio, and S. Villata (2022, 7). Fallacious argument classification in political debates. In L. D. Raedt (Ed.), Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22 , pp.\ 4143--4149. International Joint Conferences on Artificial Intelligence Organization. Main Track

2022
[2]

Hannemann, C

Habernal, I., R. Hannemann, C. Pollak, C. Klamm, P. Pauli, and I. Gurevych (2017, September). A rgotario: Computational argumentation meets serious games. In L. Specia, M. Post, and M. Paul (Eds.), Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , Copenhagen, Denmark, pp.\ 7--12. Association fo...

2017
[3]

(2022, May)

Ifan, M. (2022, May). Language as a weapon: The logical fallacies in political discourse. Journal of Society Innovation and Development\/ 3\/ (2), 18–27

2022
[4]

Liang, X., G. Tu, J. Du, and R. Xu (2024). Multi-modal attentive prompt learning for few-shot emotion recognition in conversations. Journal of Artificial Intelligence Research\/ 79 , 825--863

2024
[5]

Lotfian, R. and C. Busso (2017). Building naturalistic emotionally balanced speech corpus by retrieving emotional speech from existing podcast recordings. IEEE Transactions on Affective Computing\/ 10\/ (4), 471--483

2017
[6]

Ruggeri, and P

Mancini, E., F. Ruggeri, and P. Torroni (2024a, March). Multimodal fallacy classification in political debates. In Y. Graham and M. Purver (Eds.), Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers) , St. Julian ' s, Malta, pp.\ 170--178. Association for Computational Linguistics
[7]

Ruggeri, and P

Mancini, E., F. Ruggeri, and P. Torroni (2024b). Multimodal fallacy classification in political debates. pp.\ 170--178. Association for Computational Linguistics
[8]

Mestre, R., S. E. Middleton, M. Ryan, M. Gheasi, T. Norman, and J. Zhu (2023). Augmenting pre-trained language models with audio feature embedding for argumentation mining in political debates. In Findings of the Association for Computational Linguistics: EACL 2023 , pp.\ 274--288

2023

Show all 16 references

[9]

Mouchel, L., D. Paul, S. Cui, R. West, A. Bosselut, and B. Faltings (2025). A logical fallacy-informed framework for argument generation

2025
[10]

Musi, E. and C. Reed (2022). From fallacies to semi-fake news: Improving the identification of misinformation triggers across digital media. Discourse & Society\/ 33\/ (3), 349--370

2022
[11]

Tiwari, T

Tomar, M., A. Tiwari, T. Saha, and S. Saha (2023). Your tone speaks louder than your face! modality order infused multi-modal sarcasm detection. In Proceedings of the 31st ACM International Conference on Multimedia , pp.\ 3926--3933

2023
[12]

Van Eemeren, F. H. and R. Grootendorst (1987, September). Fallacies in pragma-dialectical perspective. Argumentation\/ 1\/ (3), 283--301

1987
[13]

Wagemans, J. H. (2023, January). How to identify an argument type? On the hermeneutics of persuasive discourse. Journal of Pragmatics\/ 203 , 117--129

2023
[14]

Triantafyllopoulos, H

Wagner, J., A. Triantafyllopoulos, H. Wierstorf, M. Schmitt, F. Burkhardt, F. Eyben, and B. W. Schuller (2023). Dawn of the transformer era in speech emotion recognition: closing the valence gap. IEEE Transactions on Pattern Analysis and Machine Intelligence\/ 45\/ (9), 10745--10759

2023
[15]

Zhang, and P

Wu, Y., S. Zhang, and P. Li (2025). Multi-modal emotion recognition in conversation based on prompt learning with text-audio fusion features. Scientific Reports\/ 15\/ (1), 8855

2025
[16]

Zhao, J., R. Li, Q. Jin, X. Wang, and H. Li (2022). Memobert: Pre-training model with prompt-based learning for multimodal emotion recognition. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp.\ 4703--4707. IEEE

2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Haddadan, V

Goffredo, P., S. Haddadan, V. Vorakitphan, E. Cabrio, and S. Villata (2022, 7). Fallacious argument classification in political debates. In L. D. Raedt (Ed.), Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22 , pp.\ 4143--4149. International Joint Conferences on Artificial Intelligence Organization. Main Track

2022

[2] [2]

Hannemann, C

Habernal, I., R. Hannemann, C. Pollak, C. Klamm, P. Pauli, and I. Gurevych (2017, September). A rgotario: Computational argumentation meets serious games. In L. Specia, M. Post, and M. Paul (Eds.), Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , Copenhagen, Denmark, pp.\ 7--12. Association fo...

2017

[3] [3]

(2022, May)

Ifan, M. (2022, May). Language as a weapon: The logical fallacies in political discourse. Journal of Society Innovation and Development\/ 3\/ (2), 18–27

2022

[4] [4]

Liang, X., G. Tu, J. Du, and R. Xu (2024). Multi-modal attentive prompt learning for few-shot emotion recognition in conversations. Journal of Artificial Intelligence Research\/ 79 , 825--863

2024

[5] [5]

Lotfian, R. and C. Busso (2017). Building naturalistic emotionally balanced speech corpus by retrieving emotional speech from existing podcast recordings. IEEE Transactions on Affective Computing\/ 10\/ (4), 471--483

2017

[6] [6]

Ruggeri, and P

Mancini, E., F. Ruggeri, and P. Torroni (2024a, March). Multimodal fallacy classification in political debates. In Y. Graham and M. Purver (Eds.), Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 2: Short Papers) , St. Julian ' s, Malta, pp.\ 170--178. Association for Computational Linguistics

[7] [7]

Ruggeri, and P

Mancini, E., F. Ruggeri, and P. Torroni (2024b). Multimodal fallacy classification in political debates. pp.\ 170--178. Association for Computational Linguistics

[8] [8]

Mestre, R., S. E. Middleton, M. Ryan, M. Gheasi, T. Norman, and J. Zhu (2023). Augmenting pre-trained language models with audio feature embedding for argumentation mining in political debates. In Findings of the Association for Computational Linguistics: EACL 2023 , pp.\ 274--288

2023