Pith. sign in

REVIEW 5 major objections 6 minor 37 references

An Adaptive Supervised Contrastive Learning Framework for Implicit Sexism Detection in Digital Social Networks

T0 review · 5 major / 6 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 Thresholded contrastive learning sharpens detection of implicit sexism.

desk verdict A reasonable kitchen-sink model for implicit sexism detection, but the central 'learned threshold' claim is contradicted by the paper's own methods, and the bilingual benchmark makes the headline numbers uninterpretable. read the letter →

arxiv 2507.05271 v1 pith:2UZLVSQJ submitted 2025-07-03 cs.CL

Mohammad Zia Ur Rehman , Aditya Shah , Nagendra Kumar This is my paper
classification cs.CL
keywords sexismdetectionimplicitsupervisedcontrastivelearninglearnablesimilaritythresholdword-levelattentionEXIST2021MLSCRoBERTa
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

ASCEND is a framework for detecting implicit sexism in social-media posts, the kind of sexist content that does not rely on slurs or obvious abuse. The paper's central claim is that adding an adaptive supervised contrastive objective to a RoBERTa classifier makes it substantially better at this task than prior specialized systems and much larger language models. Instead of treating every same-label pair as a positive training example, ASCEND only pulls together pairs whose cosine similarity exceeds a learnable threshold, which the authors argue cuts false positives and false negatives. The final prediction combines the contrastive loss with cross-entropy, word-level attention, and sentiment, emotion, and toxicity features. Reported macro-F1 gains over comparison systems average 9.86%, 29.63%, and 32.51% across binary sexism detection, sexism categorization, and multi-label categorization.

What carries the argument

The load-bearing mechanism is the adaptive supervised contrastive loss, a modified version of supervised contrastive learning in which a pair of same-label samples counts as positive only if their cosine similarity $\text{sim}(z_i, z_j)$ is above a learnable threshold; all other pairs, including same-label but low-similarity pairs, are treated as negatives. Word-level attention is applied on top of RoBERTa's last hidden states to build a weighted sentence representation, and the [CLS] representation, attention features, VADER sentiment scores, NRCLex emotion scores, and ToxicBert toxicity scores are concatenated for classification. The final objective is $\mathcal{L} = \mathcal{L}_{CL} + \mathcal{L}_{CE}$, so the contrastive loss shapes the embedding space while cross-entropy drives the actual label prediction. The threshold is the part that makes the contrastive learning adaptive, because it decides which semantic neighbors are worth pulling together.

What would settle it

Report ASCEND's macro-F1 separately on the English and Spanish subsets of EXIST 2021, and compare against a version with translation or a multilingual encoder; if Spanish-only scores collapse to near chance, the reported combined-dataset gains do not hold for the full multilingual benchmark.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a selective supervised contrastive loss makes a RoBERTa-based classifier better at implicit sexism detection than previous specialized systems and large language models. The key move is to use only same-label pairs whose cosine similarity exceeds a learned threshold as positives, jointly optimizing a contrastive loss with cross-entropy, so the model builds an embedding space where subtle sexist cues cluster tightly. The authors report macro-F1 scores of 79.05% on EXIST Task 1, 58.85% on EXIST Task 2, and 68.96% on MLSC, with average gains of 9.86%, 29.63%, and 32.51% over compared methods. Their claim is that these gains come from thresholding, attention, and the extra perception features, not from model scale.

Load-bearing premise

The system assumes the English-only models it uses for text, sentiment, emotion, and toxicity can be applied unchanged to the full EXIST 2021 dataset, even though that dataset contains Spanish tweets and no translation or language handling is described.

Editorial extensions

If this is right

  • If the central claim holds, online moderation systems can detect implicit sexism without relying on billion-parameter LLMs, since ASCEND's compact RoBERTa-based model reports higher macro-F1 than LLaMa-3.1 and GPT-3.5 on the tested benchmarks.
  • The thresholding rule implies that same-label but semantically unrelated posts will no longer be forced together in embedding space, so the model should produce fewer false positives on neutral statements and fewer false negatives on subtly sexist ones.
  • Word-level attention plus sentiment, emotion, and toxicity features gives a route to explainable cues: the attended tokens and feature scores can be inspected when a post is flagged, which matters for moderation decisions.
  • Across the three tasks (binary, categorization, multi-label), the same architecture is claimed to work without task-specific redesign, suggesting a single framework can cover detection and fine-grained tagging.

Reading between the lines

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

  • Inference: the learnable threshold can be viewed as an adaptive margin between same-label and cross-label pairs; if that interpretation is right, similar thresholded contrastive objectives should transfer to other fine-grained abusive-language tasks such as implicit hate speech or cyberbullying detection.
  • Inference: because the EXIST 2021 dataset contains Spanish tweets and all feature extractors are English-only, a natural extension is to report per-language results or to swap in a multilingual encoder; the combined macro-F1 may be dominated by the English portion.
  • Inference: the ablation sweep suggests performance is sensitive to the threshold and peaks near 0.7, so future work could test whether the learned threshold converges to a similar value or whether a fixed threshold is sufficient.
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 / 6 minor

Summary. The paper introduces ASCEND, a supervised contrastive learning framework for detecting implicit sexism in social media posts. The architecture combines RoBERTa embeddings with word-level attention, sentiment/emotion/toxicity features, and a contrastive loss that uses a cosine-similarity threshold to define positive pairs; the final objective is the sum of a cross-entropy loss and the contrastive loss. The method is evaluated on EXIST 2021 (binary sexism detection and multi-class sexism categorization) and MLSC (multi-label classification), with reported macro-F1 gains over transformer baselines, prior state-of-the-art systems, and large language models.

Significance. If the reported results were fully supported, the paper would make a useful contribution: a compact RoBERTa-based model with supervised contrastive learning and auxiliary features that outperforms much larger LLMs on three sexism-related tasks. The ablations suggest each added component contributes a modest improvement, and the qualitative analysis is informative. However, the central claim of an adaptive, learned threshold is not implemented as described, the threshold selection appears to be done on the test set, and the bilingual composition of EXIST 2021 is not addressed by the English-only feature extractors. These issues are load-bearing, and without error bars or significance tests the reported gains are not substantiated.

major comments (5)
  1. [Section 3.3.1, Eq. (5), Eq. (12), Algorithm 1, Section 4.3.3] The paper claims in Section 3.3.1 that the cosine similarity threshold "is learned by the model during training," but no differentiable threshold parameter appears in the contrastive loss (Eq. 5), the total loss (Eq. 12), or Algorithm 1. The positive-pair mask is a hard threshold on cosine similarity, which cannot be trained by gradient descent as described. Section 4.3.3 instead treats the threshold as a hyperparameter swept from 0.5 to 1.0, with 0.7 reproducing the exact ASCEND scores in Table 5. Section 4.1 describes no validation split, so this sweep appears to be performed on the test partition. Thus the reported improvements in Tables 5-7 and the abstract cannot be attributed to an adaptive mechanism; they are the result of test-set model selection.
  2. [Section 4.1.1, Sections 3.2.1 and 3.2.2] EXIST 2021 contains both English and Spanish tweets, but the feature extractors RoBERTa, VADER, NRCLex, and ToxicBert are English-only tools. The paper does not describe any translation, language filtering, or multilingual adaptation for Spanish posts. Applying English-only sentiment, emotion, and toxicity lexicons to Spanish text can produce unreliable features, so the reported macro-F1 scores on the full combined dataset cannot be taken at face value. The authors should report per-language performance or describe a valid multilingual handling procedure.
  3. [Tables 5, 6, 7] All results are reported from what appear to be single runs, with no error bars, confidence intervals, multiple seeds, or significance tests. Several headline gains are small: in Table 7 ASCEND's macro-F1 of 0.6896 exceeds the 'Heirarchy' baseline of 0.6850 by only 0.46 percentage points, and in Table 5 the gain over ROHNEIL is 0.35 percentage points. Without variance estimates, the statement that ASCEND "significantly outperforms" prior methods is unsupported.
  4. [Abstract and Section 4.2.1] The abstract's claimed average Macro-F1 improvements of 9.86%, 29.63%, and 32.51% do not match the numbers in Tables 5-7. For example, against the best baseline in Table 5 (ROHNEIL, macro-F1 0.7870) ASCEND improves by 0.35 percentage points, while against GPT-3.5 in the same table the improvement is 36.14% relative. The origin of the abstract percentages is unclear and needs correction, or the reporting basis must be stated explicitly.
  5. [Section 4.2 comparisons] Several baselines may not be directly comparable. The CLassifiers [35] and ROHNEIL [36] systems are from EXIST 2023, whereas the paper evaluates on EXIST 2021; the evaluation protocol for GPT-3.5 and LLaMa-3.1 is not described (zero-shot, few-shot, or fine-tuned). The 'Heirarchy' baseline in Table 7 is not defined as a distinct method in the reference list. These issues make the claimed superiority over prior work less reliable.
minor comments (6)
  1. [Section 4.3.2] The text says the feature ablation results are shown in Figure 2, but Figure 3 is the relevant figure; Figure 2 shows the loss and attention ablation.
  2. [Algorithm 1] In the pseudocode, the function named 'sexism detection' returns Pred, while the expected output is the predicted label y-hat; the loss function is defined but not invoked inside the main function, and the contrastive loss input is not clearly specified.
  3. [Section 3.2.1] The statement that RoBERTa is made up of an encoder and a decoder is inaccurate; RoBERTa is an encoder-only transformer.
  4. [Equations (2) and (4)] The set notation B={b_x}^X_{x=1} and L={e_x}^X_{x=1} is not well-formed for a sequence; these should be indexed families or vectors.
  5. [Section 4.2.1, Table 5] The sentence listing improvements over "current state-of-the-art approaches" gives metric order as precision, recall, accuracy, F1, but Table 5 lists accuracy first; the numbers should be matched to metrics carefully.
  6. [Table 7] The baseline name 'Heirarchy' is not defined in the reference list; if it refers to the hierarchical model of Parikh et al. [4], the citation should be made directly at that table entry.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ASCEND's reported gains are external-benchmark evaluations; the 'learned threshold' and bilingual-feature issues are validity concerns, not circular reasoning.

full rationale

The paper's derivation is an empirical pipeline: RoBERTa features, a thresholded supervised contrastive loss (Eq. 5), word-level attention (Eqs. 7-8), sentiment/emotion/toxicity features, and a cross-entropy loss (Eq. 11) combined in Eq. 12, evaluated on the external EXIST 2021 and MLSC benchmarks. The reported macro-F1 values are measurements against held-out test labels; none is obtained by algebraic substitution from the model's inputs or from a fitted parameter renamed as a prediction. The authors' earlier works are cited only for contextual support and are not load-bearing in the derivation. Two validity concerns should be kept separate from circularity: (1) Section 3.3.1 calls the threshold 'learned by the model during training,' but no threshold term or update rule appears in Eq. 5, Eq. 12, or Algorithm 1, and Section 4.3.3 instead sweeps thresholds from 0.5 to 1.0 with no validation split described, suggesting possible test-set selection; (2) Section 4.1.1 states that EXIST contains both English and Spanish tweets, while the feature extractors in Section 3.2 are English-only. Both are experimental-validity or reporting concerns, not cases where a claimed result is equivalent to its input by definition or by self-citation. The paper is therefore self-contained against external benchmarks with no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method relies on standard losses and on an unstated language-transfer assumption. The only free constant specific to the method is the positive-pair similarity threshold, which the paper describes as learned but actually tunes by sweeping. No new entities are introduced.

free parameters (2)
  • Cosine similarity threshold = 0.7 (reported best in ablation sweep)
    Controls which same-label pairs count as positives in the contrastive loss (Section 3.3.1); Section 4.3.3 selects 0.7 by sweeping, so it functions as a fitted hyperparameter despite being called learnable.
  • Contrastive temperature tau = Not reported
    Appears in Eq. 5 as the temperature that scales cosine similarities; its value is never given, yet it strongly affects the contrastive gradient.
assumptions (3)
  • standard math Supervised contrastive loss (Eq. 5) is taken from Khosla et al. [20].
    The formula is used as a building block; this is accepted prior work, not a claim of this paper.
  • standard math Cross-entropy loss (Eq. 11) is the standard classification objective.
    No novelty is claimed for the loss itself.
  • domain assumption English-only feature extractors transfer to the bilingual EXIST 2021 data.
    Sections 3.2.1 and 3.2.2 select RoBERTa, VADER, NRCLex, and ToxicBert; Section 4.1.1 says EXIST 2021 contains English and Spanish posts. No translation or language handling is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Adaptive Supervised Contrastive Learning Framework for Implicit Sexism Detection in Digital Social Networks." pith.science (2026). https://pith.science/paper/2UZLVSQJ

@misc{pith2026250705271,
  author       = {Pith},
  title        = {Pith review of: An Adaptive Supervised Contrastive Learning Framework for Implicit Sexism Detection in Digital Social Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2UZLVSQJ}},
  note         = {Machine review of arXiv:2507.05271}
}
read the original abstract

The global reach of social media has amplified the spread of hateful content, including implicit sexism, which is often overlooked by conventional detection methods. In this work, we introduce an Adaptive Supervised Contrastive lEarning framework for implicit sexism detectioN (ASCEND). A key innovation of our method is the incorporation of threshold-based contrastive learning: by computing cosine similarities between embeddings, we selectively treat only those sample pairs as positive if their similarity exceeds a learnable threshold. This mechanism refines the embedding space by robustly pulling together representations of semantically similar texts while pushing apart dissimilar ones, thus reducing false positives and negatives. The final classification is achieved by jointly optimizing a contrastive loss with a cross-entropy loss. Textual features are enhanced through a word-level attention module. Additionally, we employ sentiment, emotion, and toxicity features. Evaluations on the EXIST2021 and MLSC datasets demonstrate that ASCEND significantly outperforms existing methods, with average Macro F1 improvements of 9.86%, 29.63%, and 32.51% across multiple tasks, highlighting its efficacy in capturing the subtle cues of implicit sexist language.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 30 canonical work pages

  1. [35]

    Classifiers at exist 2023: detecting sexism in spanish and english tweets with xlm-t,

    G. Radler, B. Ersoy, and S. Carpentieri, “Classifiers at exist 2023: detecting sexism in spanish and english tweets with xlm-t,” Working Notes of CLEF , 2023

    work page 2023

  2. [36]

    Roh neil@ exist2023: Detecting sexism in tweets using multilingual language models.,

    R. Koonireddy and N. Adel, “Roh neil@ exist2023: Detecting sexism in tweets using multilingual language models.,” in CLEF (Working Notes), pp. 967–984, 2023

    work page 2023

  3. [1]

    Internet and Social Media Users in the World 2024

    A. Petrosyan, “Internet and Social Media Users in the World 2024. ” https://www.statista.com/statistics/ 617136/digital-population-worldwide/, 2024. Accessed: 2024-10-29

    work page 2024

  4. [2]

    User-aware multilingual abusive content detection in social media,

    M. Z. U. Rehman, S. Mehta, K. Singh, K. Kaushik, and N. Kumar, “User-aware multilingual abusive content detection in social media,” Information Processing & Management, vol. 60, no. 5, p. 103450, 2023

    work page 2023

  5. [3]

    Overview of exist 2021: sex- ism identification in social networks,

    F. J. Rodr´ ıguez-Sanchez, J. C. de Albornoz, L. Plaza, J. Gonzalo, P. Rosso, M. Comet, and T. Donoso, “Overview of exist 2021: sex- ism identification in social networks,” Proces. del Leng. Natural, vol. 67, pp. 195–207, 2021

    work page 2021

  6. [4]

    Multi-label Categorization of Accounts of Sexism using a Neural Framework

    P. Parikh, H. Abburi, P. Badjatiya, R. Krish- nan, N. Chhaya, M. Gupta, and V. Varma, “Multi-label categorization of accounts of sexism using a neural framework,” arXiv preprint arXiv:1910.04602, 2019

    work page Pith review arXiv 1910

  7. [5]

    An efficient bert based approach to detect aggression and misogyny,

    S. Dutta, U. Majumder, and S. K. Naskar, “An efficient bert based approach to detect aggression and misogyny,” in Proceedings of the 18th International Conference on Natural Language Processing (ICON) , pp. 493–498, 2021

    work page 2021

  8. [6]

    Aggression and misogyny detection using bert: A multi-task approach,

    N. S. Samghabadi, P. Patwa, S. Pykl, P. Mukherjee, A. Das, and T. Solorio, “Aggression and misogyny detection using bert: A multi-task approach,” in Proceedings of the second workshop on trolling, aggression and cyberbullying, pp. 126–131, 2020

    work page 2020

Show all 37 references
  1. [7]

    Roberta: A robustly opti- mized bert pretraining approach,

    Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V. Stoyanov, “Roberta: A robustly opti- mized bert pretraining approach,” 2019

    2019

  2. [8]

    Hierarchical attention-enhanced contex- tual capsulenet for multilingual hope speech detection,

    M. Z. U. Rehman, D. Raghuvanshi, H. Pachar, C. S. Raghaw, and N. Kumar, “Hierarchical attention-enhanced contex- tual capsulenet for multilingual hope speech detection,” Expert Systems with Applications, vol. 268, p. 126285, 2025

    2025

  3. [9]

    Hateful symbols or hateful people? predictive features for hate speech detection on twitter,

    Z. Waseem and D. Hovy, “Hateful symbols or hateful people? predictive features for hate speech detection on twitter,” in Proceedings of the NAACL student research workshop , pp. 88–93, 2016

    2016

  4. [10]

    Hate speech detection in the indonesian language: A dataset and prelimi- nary study,

    I. Alfina, R. Mulia, M. I. Fanany, and Y. Ekanata, “Hate speech detection in the indonesian language: A dataset and prelimi- nary study,” in 2017 international conference on advanced computer science and informa- tion systems (ICACSIS), pp. 233–238, IEEE, 2017

    2017

  5. [11]

    Deep learning for hate speech detection in tweets,

    P. Badjatiya, S. Gupta, M. Gupta, and V. Varma, “Deep learning for hate speech detection in tweets,” in Proceedings of the 26th international conference on World Wide Web companion, pp. 759–760, 2017

    2017

  6. [12]

    Detecting hate speech on twitter using a convolution-gru based deep neural network,

    Z. Zhang, D. Robinson, and J. Tepper, “Detecting hate speech on twitter using a convolution-gru based deep neural network,” in The Semantic Web: 15th International Conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, Proceedings 15 , pp. 745–760, Springer, 2018

    2018

  7. [13]

    Learning phrase representations using rnn encoder-decoder for statisti- cal machine translation,

    K. Cho, “Learning phrase representations using rnn encoder-decoder for statisti- cal machine translation,” arXiv preprint arXiv:1406.1078, 2014

    2014 arXiv

  8. [14]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  9. [15]

    Bert: Pre-training of deep bidi- rectional transformers for language under- standing,

    J. Devlin, “Bert: Pre-training of deep bidi- rectional transformers for language under- standing,” arXiv preprint arXiv:1810.04805 , 2018. 14

    2018 arXiv

  10. [16]

    A context- aware attention and graph neural network- based multimodal framework for misogyny detection,

    M. Z. U. Rehman, S. Zahoor, A. Manzoor, M. Maqbool, and N. Kumar, “A context- aware attention and graph neural network- based multimodal framework for misogyny detection,” Information Processing & Man- agement, vol. 62, no. 1, p. 103895, 2025

    2025

  11. [17]

    A contrastive topic-aware attentive frame- work with label encodings for post-disaster resource classification,

    S. S. Dar, M. K. Karandikar, M. Z. U. Rehman, S. Bansal, and N. Kumar, “A contrastive topic-aware attentive frame- work with label encodings for post-disaster resource classification,” Knowledge-Based Systems, vol. 304, p. 112526, 2024

    2024

  12. [18]

    An explainable contrastive- based dilated convolutional network with transformer for pediatric pneumonia detec- tion,

    C. S. Raghaw, P. S. Bhore, M. Z. U. Rehman, and N. Kumar, “An explainable contrastive- based dilated convolutional network with transformer for pediatric pneumonia detec- tion,” Applied Soft Computing , vol. 167, p. 112258, 2024

    2024

  13. [19]

    A simple framework for con- trastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for con- trastive learning of visual representations,” in International conference on machine learn- ing, pp. 1597–1607, PMLR, 2020

    2020

  14. [20]

    Supervised contrastive learn- ing,

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y. Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learn- ing,” Advances in neural information pro- cessing systems , vol. 33, pp. 18661–18673, 2020

    2020

  15. [21]

    Vi2clr: Video and image for visual con- trastive learning of representation,

    A. Diba, V. Sharma, R. Safdari, D. Lotfi, S. Sarfraz, R. Stiefelhagen, and L. Van Gool, “Vi2clr: Video and image for visual con- trastive learning of representation,” in Pro- ceedings of the IEEE/CVF international con- ference on computer vision , pp. 1502–1512, 2021

    2021

  16. [22]

    Bootstrap your own latent-a new approach to self-supervised learning,

    J.-B. Grill, F. Strub, F. Altch´ e, C. Tallec, P. Richemond, E. Buchatskaya, C. Doer- sch, B. Avila Pires, Z. Guo, M. Ghesh- laghi Azar, et al. , “Bootstrap your own latent-a new approach to self-supervised learning,” Advances in neural information processing systems, vol. 33,...

    2020

  17. [23]

    Simcse: Sim- ple contrastive learning of sentence embed- dings,

    T. Gao, X. Yao, and D. Chen, “Simcse: Sim- ple contrastive learning of sentence embed- dings,” arXiv preprint arXiv:2104.08821 , 2021

    2021 arXiv

  18. [24]

    Multilingual sexism identification using contrastive learning.,

    J. Angel, S. T. Aroyehun, and A. F. Gelbukh, “Multilingual sexism identification using contrastive learning.,” in CLEF (Work- ing Notes), pp. 855–861, 2023

    2023

  19. [25]

    Hate speech detection via dual contrastive learning,

    J. Lu, H. Lin, X. Zhang, Z. Li, T. Zhang, L. Zong, F. Ma, and B. Xu, “Hate speech detection via dual contrastive learning,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023

    2023

  20. [26]

    Kisanqrs: A deep learning- based automated query-response system for agricultural decision-making,

    M. Z. U. Rehman, D. Raghuvanshi, and N. Kumar, “Kisanqrs: A deep learning- based automated query-response system for agricultural decision-making,” Comput- ers and Electronics in Agriculture , vol. 213, p. 108180, 2023

    2023

  21. [27]

    A hybrid filtering for micro-video hashtag recommen- dation using graph-based deep neural net- work,

    S. Bansal, K. Gowda, M. Z. U. Rehman, C. S. Raghaw, and N. Kumar, “A hybrid filtering for micro-video hashtag recommen- dation using graph-based deep neural net- work,” Engineering Applications of Artificial Intelligence, vol. 138, p. 109417, 2024

    2024

  22. [28]

    Vader: A parsimo- nious rule-based model for sentiment analysis of social media text,

    C. Hutto and E. Gilbert, “Vader: A parsimo- nious rule-based model for sentiment analysis of social media text,” in Proceedings of the international AAAI conference on web and social media, vol. 8, pp. 216–225, 2014

    2014

  23. [29]

    Crowd- sourcing a word–emotion association lexi- con,

    S. M. Mohammad and P. D. Turney, “Crowd- sourcing a word–emotion association lexi- con,” Computational intelligence , vol. 29, no. 3, pp. 436–465, 2013

    2013

  24. [30]

    Hierarchical attention net- works for document classification,

    Z. Yang, D. Yang, C. Dyer, X. He, A. Smola, and E. Hovy, “Hierarchical attention net- works for document classification,” in Pro- ceedings of the 2016 conference of the North American chapter of the association for com- putational linguistics: human language tech- nologies, pp...

    2016

  25. [31]

    Unsupervised cross-lingual representation learning at scale,

    A. Conneau, “Unsupervised cross-lingual representation learning at scale,” arXiv preprint arXiv:1911.02116, 2019

    1911 arXiv

  26. [32]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” Journal of machine learning research, vol. 21, no. 140, pp. 1–67, 2020

    2020

  27. [33]

    The llama 3 herd of models,

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fan, et al. , “The llama 3 herd of models,” arXiv preprint arXiv:2407.21783, 2024. 15

    2024 arXiv

  28. [34]

    Bert-based ensemble learning for multi- aspect hate speech detection,

    A. C. Mazari, N. Boudoukhani, and A. Djef- fal, “Bert-based ensemble learning for multi- aspect hate speech detection,” Cluster Com- puting, vol. 27, no. 1, pp. 325–339, 2024

    2024

  29. [37]

    A hybrid deep bilstm-cnn for hate speech detection in multi-social media,

    A. Kumar, S. Kumar, K. Passi, and A. Mahanti, “A hybrid deep bilstm-cnn for hate speech detection in multi-social media,” ACM Transactions on Asian and Low-Resource Language Information Pro- cessing, 2024. 16

    2024

Pith tools

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