Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

CHUCKLE claims that ordering speech-emotion training clips by how hard human annotators found them beats both no-curriculum and data-driven baselines in accuracy and training speed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 10:26 UTC pith:KF7PHI2J

load-bearing objection A useful ordering heuristic for SER curricula with one real positive result, but the perception-based mechanism is asserted rather than shown. the 5 major comments →

arxiv 2510.09382 v3 pith:KF7PHI2J submitted 2025-10-10 cs.LG

CHUCKLE -- When Humans Teach AI To Learn Emotions The Easy Way

Ankush Pratap Singh , Houwei Cao , Yong Liu This is my paper
classification cs.LG
keywords curriculum learningspeech emotion recognitionhuman perceptionannotator agreementintended labelCREMA-Dtraining efficiency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes CHUCKLE, a curriculum-learning method for speech emotion recognition that defines sample difficulty from crowd-sourced human annotations instead of heuristic, data-driven, or model-based scores. It sorts clips into four difficulty bins using two signals: how much annotators agree with each other, and whether their perceived emotion matches the actor's intended emotion. Under the assumption that clips humans find hard are also hard for neural networks, training starts on clear, well-aligned clips and adds harder bins later. On the CREMA-D dataset, the best rule-based curriculum improves mean macro accuracy by 6.56% relative for LSTMs and 1.61% relative for Transformers, while also reducing the number of gradient updates needed. The practical appeal is that annotator disagreement, usually treated as label noise, becomes a free and interpretable training signal.

Core claim

The paper's central claim is that the structure of human perception, specifically how consistently raters agree on an acted emotion and whether that agreement aligns with the intended label, is a superior difficulty signal for curriculum learning in speech emotion recognition. Clips are grouped into four categories: Clear Match, Clear Mismatch, Ambiguous Match, and Ambiguous Mismatch. Curricula that order these categories by prioritizing agreement strength first, most notably the Intended-Perceived Agreement 1 ordering (Clear Match, Clear Mismatch, Ambiguous Match, Ambiguous Mismatch), consistently outperform both non-curriculum training and data-driven curricula such as entropy or intended-

What carries the argument

The central object is the Intended-Perceived Agreement curriculum, a rule-based difficulty ladder built from annotator behavior. It assigns each clip to one of four named categories based on whether the perceived-label majority is clear or split and whether that majority matches the intended emotion: Clear Match, Clear Mismatch, Ambiguous Match, and Ambiguous Mismatch. The machinery works by converting raw annotator disagreement and misalignment, which are normally discarded as noise, into a training order that progresses from unambiguous and aligned clips to misleading or ambiguous ones.

Load-bearing premise

The load-bearing premise is that clips human annotators find ambiguous or misleading are also the clips machine models find hard; if human and model difficulty rankings do not correlate, the perception-based ordering has no reason to help and could even mislead training.

What would settle it

A direct test: train the same two models with the CHUCKLE bin order reversed, hardest-first, and with a shuffled bin order, then compare mean macro accuracy and convergence speed; the paper's claim predicts a clear drop in both. A second check is to compute, on a held-out dataset, how well each clip's bin predicts per-sample loss under ordinary non-curriculum training; near-zero or negative correlation would indicate that human-perceived difficulty does not transfer to model difficulty.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If CHUCKLE is correct, crowd-sourced emotion datasets with per-item annotator labels gain extra value: their disagreement structure becomes a training signal rather than just noise to be aggregated away.
  • Human-interpretable rule-based curricula can outperform continuous data-driven difficulty scores such as entropy or intended-emotion proportion, at least for acted speech emotion recognition.
  • Curriculum training can deliver comparable or better accuracy with fewer gradient updates, lowering the compute cost of training speech emotion models.
  • Most of the accuracy gain comes between the Easy and Borderline Easy stages, which together contain most of the data, suggesting that early staging matters more than the ordering of the hardest bins.
  • Transformers are more sensitive to the distribution shift between curriculum stages, but the best rule-based curriculum still improves them, whereas random or entropy-based curricula can hurt them.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's core assumption could be tested directly by comparing CHUCKLE's bin order with model-derived difficulty, such as per-sample loss or gradient norms under ordinary training; if the two rankings diverge, a curriculum that combines human and model difficulty might do better than either alone.
  • A natural extension the authors leave implicit is to apply the same binning to video or multimodal data, where agreement rates differ sharply by emotion (for example, audio sadness agreement is only 16.4% but video agreement is 33.4%); the optimal curriculum order may change by modality.
  • Because CHUCKLE uses only four coarse bins, a finer-grained ranking within each bin by entropy or alignment could smooth the easy-to-hard transition and might reduce the Transformer's temporary instability in stage 2.
  • The efficiency claim invites a broader cost comparison: measuring wall-clock time and energy in addition to gradient updates would show whether the staged schedule remains cheaper after the overhead of binning and multiple training stages is included.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes CHUCKLE, a curriculum-learning framework for speech emotion recognition that orders training samples by human-perceived difficulty, derived from annotator agreement and alignment in the CREMA-D dataset. Three rule-based curricula (IPA1, IPA2, IPA3) and two data-driven curricula (intended-emotion score, entropy score) are compared against non-curriculum and random-order baselines on LSTM and Transformer classifiers. The authors report that the IPA1 curriculum improves mean macro accuracy by 6.56% for LSTMs and 1.61% for Transformers relative to the non-curriculum baseline, while reducing the number of gradient updates. The paper claims rule-based curricula consistently outperform baselines and that results hold in both subject-dependent and subject-independent settings.

Significance. If the results are robust, the paper offers a simple, interpretable mechanism for injecting human perception into curriculum design for subjective tasks like emotion recognition, where label noise is intrinsic. The use of public annotation metadata (agreement, intended-perceived alignment) is a plausible and practical idea, and the efficiency analysis is a useful addition. However, the central claim of consistent rule-based superiority is weakened by the reported numbers themselves, the core human-to-model difficulty transfer assumption is not tested, and an abstract claim about subject-independent evaluation is contradicted by the experimental setup. The paper's value would be substantially increased by addressing these points; as it stands, the evidence supports only a more modest claim about one particular curriculum (IPA1) on this dataset.

major comments (5)
  1. [§4, Table 4; §1 Contributions] The claim that rule-based curricula 'consistently outperform' non-curriculum and data-driven curricula is not supported by Table 4. For Transformers, IPA2 (0.5861) and IPA3 (0.5872) are both below the non-curriculum baseline (0.5946), and Random Curriculum (0.5212) is far worse. Only IPA1 shows a statistically significant improvement. For LSTMs, all curricula improve, but only IPA1 and IPA3 are significant. The conclusion should be limited to IPA1 or otherwise qualified.
  2. [Abstract; §4, Table 3] The abstract states results are obtained 'in both subject-dependent and subject-independent settings,' but the experimental section only describes an 80/20 train-test split 'per subject, per intended emotion,' which is a subject-dependent split. No subject-independent (e.g., leave-speakers-out) experiment appears anywhere in the manuscript. This claim must be removed or supported by the missing experiments.
  3. [Abstract; §1; §3.1] The core assumption that clips difficult for human annotators are similarly difficult for neural networks is never tested. No correlation is reported between human difficulty measures (entropy, intended-emotion proportion, agreement category) and model per-sample loss or error. The observed benefit of IPA1 could stem from other properties of the ordering (e.g., the large Clear Mismatch bin being placed early). Add an analysis, for example Spearman correlation between human difficulty scores and per-sample loss of a baseline non-curriculum model, or bin-wise accuracy of a model trained without curriculum.
  4. [§3.1.2, Table 2] The definition of 'clear majority' is never specified, yet it determines the four rule-based categories and therefore the entire curriculum ordering. Without this threshold, the experiments are not reproducible and the threshold remains an unexamined free parameter. Please define the criterion and, ideally, test sensitivity to its value.
  5. [§4, statistical validation] The paper performs one-sided paired t-tests for each curriculum against the non-curriculum baseline without any correction for multiple comparisons. With six curricula tested per model, the chance of at least one false positive is inflated. Report adjusted p-values or use a multiple-comparison procedure. Also, selecting IPA1 after inspecting test-set results is a form of test-set fitting; a validation split or pre-registration of the primary curriculum would strengthen the claim.
minor comments (6)
  1. [Abstract; §4] The phrase 'reducing the number of gradient updates' is accurate only under the specific epoch-per-stage allocation chosen. The total number of epochs is the same as the baseline (200 for LSTM, 400 for Transformer); the reduction comes from using growing subsets. Please clarify that the comparison holds for this fixed epoch budget.
  2. [§3.2, Eq. (2)] The notation S_i is not defined explicitly as the cumulative union of bins up to stage i. As written, the formula could be misread as disjoint sets. Define S_i and note that S_1 ⊂ S_2 ⊂ ... ⊂ S_K.
  3. [Table 4] Reporting only asterisks for significance is insufficient; include p-values or confidence intervals, especially because the standard deviations are relatively large for the LSTM.
  4. [§2, Table 1] Table 1 reports multimodal match rates, but the experiments use only acoustic features. Clarify that multimodal agreement is shown for dataset characterization and is not used in the training features.
  5. [§3.1.1] The labels 'data-driven' for the entropy and intended-emotion-score curricula are confusing because these scores also come from human annotations. Consider renaming them 'continuous-score curricula' or explicitly stating that both families are perception-derived.
  6. [Figure 2] Panels (a) and (c) show training loss for a single trial. State whether the displayed trial is representative, or plot the mean over the 10 trials with error bands.

Circularity Check

0 steps flagged

No significant circularity: curricula derived from annotation statistics are evaluated on held-out accuracy; the human-to-model difficulty link is an untested assumption, not a circular reduction.

full rationale

The paper's claims are empirical rather than derivational, and no quantity is shown to be its own input. The difficulty scores in Section 3.1.1 (intended-emotion proportion and entropy) and the rule-based bins in Section 3.1.2 are computed directly from crowd-sourced annotation statistics, and are used only as training-order signals. The reported improvements (Table 4) are measured on a held-out 20% test split against intended-label macro accuracy, not on the annotation agreement statistics used to construct the curricula, so no fitted parameter is renamed as a prediction and no accuracy result is forced by the ordering rule. The stated premise that 'clips challenging for humans are similarly hard for machine learning models' is explicitly an assumption in the abstract and Section 1, not a derived conclusion; lack of correlation evidence is a correctness/validity weakness, not circularity. The only self-citation of note is the CREMA-D dataset [16], which is an external, independently reproducible benchmark and is not load-bearing in a circular sense. The practice of comparing several curricula and highlighting the best performs empirical model selection, which raises multiple-testing concerns but does not constitute equation-level circularity. Thus the paper is self-contained in its experimental logic, and no circular step can be quoted with a specific reduction.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 0 invented entities

The paper introduces no new theoretical entities but depends on the human-model difficulty correlation assumption and several hand-chosen design hyperparameters. The main parameter-selection risk is the post hoc choice of IPA1 as best.

free parameters (5)
  • number_of_curriculum_stages = 4
    The number of difficulty bins used in every curriculum is set to 4 (Easy, Borderline Easy, Borderline Tough, Tough).
  • quantile_bin_boundaries = quartiles of the difficulty score
    Samples divided into four equal-size quartiles based on difficulty scores; the cut points are data-dependent.
  • clear_majority_threshold = unspecified
    The rule-based categories in Section 3.1.2 require a 'clear majority' of perceived emotions, but the exact threshold is never defined.
  • epochs_per_stage = 50 (LSTM), 100 (Transformer)
    Each of the 4 stages trains for 50/100 epochs, regardless of subset size.
  • learning_rate_decay_range = 5e-4 to 5e-5
    Cosine annealing schedule with specified range, reset per bin.
axioms (3)
  • domain assumption Human-perceived difficulty correlates with model difficulty
    Stated explicitly in abstract; central to the method's rationale but not validated.
  • domain assumption Inter-annotator disagreement and intended-perceived alignment measure sample difficulty
    Used to define all curricula; not independently validated.
  • domain assumption Intended emotion labels are the correct targets for SER
    The model is trained to predict intended labels; ambiguity in human perception is treated as noise.

pith-pipeline@v1.3.0-alltime-deepseek · 6712 in / 11997 out tokens · 109627 ms · 2026-08-04T10:26:39.991971+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of CHUCKLE -- When Humans Teach AI To Learn Emotions The Easy Way." pith.science (2026). https://pith.science/paper/KF7PHI2J

@misc{pith2026251009382,
  author       = {Pith},
  title        = {Pith review of: CHUCKLE -- When Humans Teach AI To Learn Emotions The Easy Way},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KF7PHI2J}},
  note         = {Machine review of arXiv:2510.09382}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Curriculum learning (CL) structures training from simple to complex samples, facilitating progressive learning. However, existing CL approaches for emotion recognition often rely on heuristic, data-driven, or model-based definitions of sample difficulty, neglecting the difficulty for human perception, a critical factor in subjective tasks like emotion recognition. We propose CHUCKLE (Crowdsourced Human Understanding Curriculum for Knowledge Led Emotion Recognition), a perception-driven CL framework that leverages annotator agreement and alignment in crowd-sourced datasets to define sample difficulty, under the assumption that clips challenging for humans are similarly hard for neural networks. Experimental results suggest that CHUCKLE enhances the performance of LSTMs and Transformers over non-curriculum baselines, while reducing the number of gradient updates, thereby enhancing both training efficiency and model robustness in both subject-dependent and subject-independent settings.

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. SPICE: Synergy and Partial Information Based Curriculum Evolution

    cs.LG 2026-06 conditional novelty 5.0

    A dynamic curriculum that sorts multimodal samples by heuristic redundancy/unique/synergy scores computed from the model's own predictions improves results over static and prior dynamic curricula on four benchmarks.

Reference graph

Works this paper leans on

30 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [1]

    INTRODUCTION Emotions shape human experience, influencing communi- cation, decision-making, and social interaction. Automatic emotion recognition seeks to infer human affective states from multi-modal signals such as speech [1, 2, 3], text [3], facial expressions [3, 4], gestures [5, 6], and physiological signals [7, 8]. Among these, speech emotion recogn...

  2. [2]

    Our rule-based curricula derived from human percep- tion difficulty consistently outperform non-curriculum and data-driven curricula

  3. [3]

    We develop a novel perception-driven CL framework that integrates rule-based and data-driven curricula

  4. [4]

    To overcome this, we trained on pre-extracted acoustic features (Section 2) instead of raw audio, allowing for more effi- cient learning

    EXPERIMENTAL EV ALUA TION Deep neural networks typically require large datasets and extensive training, but our work faced limitations due to the limited number of training clips and computing resources. To overcome this, we trained on pre-extracted acoustic features (Section 2) instead of raw audio, allowing for more effi- cient learning. Training was pe...

  5. [5]

    We demonstrate that CL model training is more effi- cient, and converges faster to strong performance with fewer gradient updates

  6. [6]

    DA TASET AND FEA TURES We used the CREMA-D dataset [16], a standard audiovisual benchmark comprising 7,442 clips (≈12 hours) from 91 ac- tors, who express six emotions across 12 sentences. Each clip has one intended label and multiple perceived labels (8–12 ratings) from 2,443 raters, yielding four types of labels per clip (three perceived, one per modali...

  7. [7]

    confidently incorrect

    METHODOLOGY 3.1. Curriculum Design The design of curricula for SER must account for the subjec- tive and often ambiguous nature of emotional labels. In acted datasets such as CREMA-D, the agreement between intended and perceived labels shows how consistently an expression is recognized, while disagreement indicates ambiguity or possi- ble misinterpretatio...

  8. [8]

    Research progress of eeg-based emotion recognition: A survey,

    Y . Wang, B. Zhang, and L. Di, “Research progress of eeg-based emotion recognition: A survey,”ACM Com- put. Surv., vol. 56, no. 11, July 2024

    2024

  9. [9]

    CONCLUSION AND FUTURE WORK This study highlights the effectiveness of curriculum learn- ing for speech emotion recognition. Rule-based curricula derived from agreement and alignment of human perception consistently outperformed non-curriculum and data-driven curricula, improving both accuracy and efficiency. LSTMs achieved a 6.56% relative gain in mean ma...

  10. [10]

    A review on speech emo- tion recognition: A survey, recent advances, challenges, and the influence of noise,

    S.M. George and P.M. Ilyas, “A review on speech emo- tion recognition: A survey, recent advances, challenges, and the influence of noise,”Neurocomputing, vol. 568, pp. 127015, 2024

    2024

  11. [11]

    Speech emotion recognition: two decades in a nutshell, benchmarks, and ongoing trends,

    B. W. Schuller, “Speech emotion recognition: two decades in a nutshell, benchmarks, and ongoing trends,” Communications of the ACM, vol. 61, no. 5, pp. 90–99, Apr. 2018

    2018

  12. [12]

    A survey of deep learning-based multimodal emotion recognition: Speech, text, and face,

    H. Lian, C. Lu, S. Li, Y . Zhao, C. Tang, and Y . Zong, “A survey of deep learning-based multimodal emotion recognition: Speech, text, and face,”Entropy, vol. 25, no. 10, pp. 1440, 2023

    2023

  13. [13]

    A survey on facial emotion recognition techniques: A state-of-the-art literature review,

    F. Z. Canal, T. R. M ¨uller, J. C. Matias, G. G. Scotton, A. R. de Sa Junior, E. Pozzebon, and A. C. Sobieranski, “A survey on facial emotion recognition techniques: A state-of-the-art literature review,”Information Sciences, vol. 582, pp. 593–617, 2022

    2022

  14. [14]

    Survey on emotional body gesture recognition,

    F. Noroozi, C.A. Corneanu, D. Kami ´nska, T. Sapi ´nski, S. Escalera, and G. Anbarjafari, “Survey on emotional body gesture recognition,”IEEE Transactions on Affec- tive Computing, vol. 12, no. 2, pp. 505–523, 2021

    2021

  15. [15]

    They used quantitative disagreement measures such as entropy and error rate to rank samples

    defined perception difficulty in SER as inter-annotator disagreement. They used quantitative disagreement measures such as entropy and error rate to rank samples. In this paper, we propose CHUCKLE, a novel human perception-centered CL framework for SER that integrates data-driven strategies (entropy, proportion of intended- emotion votes) (Section 3.1.1) ...

    Pith/arXiv arXiv 2025

  16. [16]

    Comprehensive survey on recognition of emotions from body gestures,

    R. Gandi, A.Geetha, and B.R. Reddy, “Comprehensive survey on recognition of emotions from body gestures,” Journal of Informatics Education and Research, vol. 5, 01 2025

    2025

  17. [17]

    Review of studies on emotion recog- nition and judgment based on physiological signals,

    W. Lin and C. Li, “Review of studies on emotion recog- nition and judgment based on physiological signals,” Applied Sciences, vol. 13, no. 4, 2023

    2023

  18. [18]

    Curriculum learning,

    Y . Bengio, J. Louradour, R. Collobert, and J. Weston, “Curriculum learning,” inProceedings of the 26th An- nual International Conference on Machine Learning, Montreal, QC, Canada, 2009, pp. 41–48

    2009

  19. [19]

    A curriculum learn- ing method for improved noise robustness in automatic speech recognition,

    S. Braun, D. Neil, and S.C. Liu, “A curriculum learn- ing method for improved noise robustness in automatic speech recognition,” in2017 25th European Signal Pro- cessing Conference (EUSIPCO), 2017, pp. 548–552

    2017

  20. [20]

    Curriculum learning based approaches for noise robust speaker recognition,

    S. Ranjan and J.H.L. Hansen, “Curriculum learning based approaches for noise robust speaker recognition,” IEEE/ACM Trans. Audio, Speech and Lang. Proc., vol. 26, no. 1, pp. 197–210, Jan. 2018

    2018

  21. [21]

    Inferring emotion from large-scale internet voice data: A semi- supervised curriculum augmentation based deep learn- ing approach,

    S. Zhou, J. Jia, Z. Wu, Z. Yang, Y . Wang, W. Chen, F. Meng, S. Huang, J. Shen, and X. Wang, “Inferring emotion from large-scale internet voice data: A semi- supervised curriculum augmentation based deep learn- ing approach,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 7, pp. 6039–6047, May 2021

    2021

  22. [22]

    Hybrid curricu- lum learning for emotion recognition in conversation,

    L. Yang, Y . Shen, Y . Mao, and L. Cai, “Hybrid curricu- lum learning for emotion recognition in conversation,” Proceedings of the AAAI Conference on Artificial Intel- ligence, vol. 36, no. 10, pp. 11595–11603, Jun. 2022

    2022

  23. [23]

    An interpretable deep mutual information curriculum metric for a robust and generalized speech emotion recognition system,

    W.C. Lin, K. Sridhar, and C. Busso, “An interpretable deep mutual information curriculum metric for a robust and generalized speech emotion recognition system,” IEEE/ACM Trans. Audio, Speech and Lang. Proc., vol. 32, pp. 5117–5130, Nov. 2024

    2024

  24. [24]

    Curriculum learning for speech emotion recognition from crowdsourced labels,

    R. Lotfian and C. Busso, “Curriculum learning for speech emotion recognition from crowdsourced labels,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 27, no. 4, pp. 815–826, 2019

    2019

  25. [25]

    Crema-d: Crowd-sourced emotional multimodal actors dataset,

    H. Cao, D. G. Cooper, M. K. Keutmann, R. C. Gur, A. Nenkova, and R. Verma, “Crema-d: Crowd-sourced emotional multimodal actors dataset,”IEEE Transac- tions on Affective Computing, vol. 5, no. 4, pp. 377–390, 2014

    2014

  26. [26]

    The in- terspeech 2013 computational paralinguistics challenge: social signals, conflict, emotion, autism,

    B. Schuller, S. Steidl, A. Batliner, A. Vinciarelli, K. Scherer, F. Ringeval, M. Chetouani, F. Weninger, F. Eyben, E. Marchi, M. Mortillaro, H. Salamin, A. Polychroniou, F. Valente, and S. Kim, “The in- terspeech 2013 computational paralinguistics challenge: social signals, conflict, emotion, autism,” inInterspeech 2013, 2013, pp. 148–152

    2013

  27. [27]

    A mathematical theory of communica- tion,

    C.E. Shannon, “A mathematical theory of communica- tion,”The Bell System Technical Journal, vol. 27, no. 3, pp. 379–423, 1948

    1948

  28. [28]

    Adam: A method for stochas- tic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochas- tic optimization,” in3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015

    2015

  29. [29]

    SGDR: stochastic gradi- ent descent with warm restarts,

    I. Loshchilov and F. Hutter, “SGDR: stochastic gradi- ent descent with warm restarts,” in5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings, 2017

    2017

  30. [30]

    Un- derstanding the difficulty of training transformers,

    L. Liu, X. Liu, J. Gao, W. Chen, and J. Han, “Un- derstanding the difficulty of training transformers,” in Proceedings of the 2020 Conference on Empirical Meth- ods in Natural Language Processing (EMNLP), Online, Nov. 2020, pp. 5747–5763, Association for Computa- tional Linguistics. 5

    2020