Pith. sign in

REVIEW 3 major objections 6 minor 54 references

Multi-task Learning with Active Learning for Arabic Offensive Speech Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper shows that jointly fine-tuning a shared Arabic transformer on offensive, violent, and vulgar speech, while actively selecting the most uncertain tweets, reaches a macro F1 of 85.42% on the OSACT2022 test set using only about…

desk verdict A plausible but unproven SOTA claim: the 85.42 is the best cell of a 144-run grid at one seed, and the 'fewer samples' claim rests on an unclear counting convention. read the letter →

arxiv 2506.02753 v1 pith:IR4EUWJ3 submitted 2025-06-03 cs.CL

Aisha Alansari , Hamzah Luqman This is my paper · ORCID
classification cs.CL
keywords Arabicoffensivespeechdetectionmulti-tasklearningactiveuncertaintysamplingemojiweightingOSACT2022AraBERTmacroF1
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 aims to establish that Arabic offensive speech detection can be made more accurate and far more data-efficient by solving three related tasks together and letting the model choose its own training examples. The proposed framework fine-tunes one shared AraBERT-TwitterV2 encoder on offensive, violent, and vulgar speech with task-weighted losses, selects the most informative tweets through entropy-based uncertainty sampling, and gives extra weight to emojis associated with offensive content. On the OSACT2022 test set, the best configuration reports a macro F1-score of 85.42%, which the authors present as state of the art and which is reached with roughly 3,336 of the 8,557 available training samples. If correct, the result matters because labeled Arabic social-media data is scarce and costly, so a pipeline that needs less data while jointly modeling three abuse categories is directly relevant to content moderation in resource-constrained settings.

What carries the argument

The carrying mechanism is the combination of hard parameter sharing and entropy-based acquisition. All three tasks share one AraBERT-TwitterV2 encoder (a BERT-style Arabic transformer pretrained on Twitter text), with a task-specific classification layer per task; the shared encoder is trained by back-propagating a weighted sum of the three task losses. Dynamic loss weighting adapts the task weights during training so high-loss tasks get more influence, and the dynamic uncertainty formula does the same for sample selection, with the offensive task initially weighted highest. Weighted emoji handling is the third component: a predefined list of emojis commonly tied to offensive speech is assigned extra weight in the input representation. The combination of these components is what the paper credits for the 85.42% result.

What would settle it

Run the same three-task MTL and uncertainty-sampling grid on OSACT2022 across, say, ten random seeds and compare mean macro F1: if the averaged best-configuration score falls below the strongest prior result in Table 5, or if fine-tuning on all 8,557 samples with the same model matches or beats the 3,336-sample configuration, the data-efficiency claim fails. A second check is to recompute the comparison using the complete OSACT2022 leaderboard; if a system not listed in Table 5 already reports a higher macro F1, the state-of-the-art claim fails.

Watch

Extended reading notes

Core claim

The central discovery is that multi-task learning and active learning reinforce each other in this setting. A single AraBERT-TwitterV2 model is fine-tuned with three classification heads—offensive, violent, and vulgar—sharing the encoder, and the total loss is a weighted sum of the three task losses under equal, static, or dynamic weighting. Active learning then scores each unlabeled tweet by the entropy of the model's predictions on all three tasks, combined through equal, weighted, or dynamic formulas, and the most uncertain tweets are selected for fine-tuning. On OSACT2022, the best configuration (dynamic loss weighting, equal uncertainty combination, weighted emojis, and ten selected samples per batch) reaches 85.42% macro F1, surpassing the 85.31% of the same model trained single-task on all 8,557 samples and the 85.17% of the strongest prior system listed in Table 5, while training the MTL model on about 3,336 samples.

Load-bearing premise

The load-bearing premise is that the best configuration found by scanning the reported grid of task weights, uncertainty formulas, sample counts, and emoji settings—evaluated on one fixed test split with a single random seed (42)—yields a stable 85.42% macro F1, and that Table 5 lists every relevant prior system for the state-of-the-art comparison.

Editorial extensions

If this is right

  • On OSACT2022, the reported 85.42% macro F1 exceeds every prior system in Table 5, including the 85.17% of the strongest listed system, so the framework would be the new reference point for Arabic offensive speech detection on this benchmark.
  • The data-efficiency target is concrete: the best config uses about 3,336 training samples instead of 8,557, so the same model can be fine-tuned with roughly 39% of the labeled corpus.
  • Auxiliary tasks help the main task: in the reported ablations, MTL configurations generally outperform the STL model at the same or smaller sample budgets, and dynamic loss weighting beats equal weighting.
  • Weighted emoji handling yields consistent gains over no-emoji or unweighted-emoji settings across most of the MTL and uncertainty combinations in Table 6.
  • Over-weighting the main task is harmful: in Table 8, raising the offensive-task loss weight from 0.70 to 0.80 decreases performance, so there is a tunable balance between the main task and its auxiliary tasks.

Reading between the lines

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

  • Inference: the active-learning gain is probably task-dependent; a direct extension would be to run the same three-task entropy sampling on another Arabic abuse benchmark (for example OSACT2020) and check whether the 3,336-sample advantage persists.
  • Inference: the sampling strategy and the loss weighting strategy are tuned separately, but the dynamic uncertainty formula already reacts to task F1 during training; a natural next step is to make the sample-selection weights co-adapt with the loss weights within a single loop.
  • Inference: because the emoji list is fixed, the framework leaves room for data-driven emoji weighting or treating emoji sequences as tokens, which could reduce the misclassifications the paper itself reports for tweets where offensive and non-offensive posts share the same emojis.
  • Inference: the headline margin over the prior best (85.42 vs 85.17) is under 0.3 points and the experiments use one seed; averaging the same grid over several seeds would show whether the advantage is stable, which is the most direct stress test of the state-of-the-art claim.
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 / 6 minor

Summary. The paper proposes a framework that combines multi-task learning (MTL) over three Arabic speech tasks (offensive, vulgar, violent) with entropy-based active learning and weighted emoji handling, evaluated on the OSACT2022 dataset. The authors report a macro F1 of 85.42%, which they claim is state-of-the-art and achieved with significantly fewer fine-tuning samples than prior work. The paper includes ablations over task-weighting strategies (equal, static, dynamic), uncertainty-sampling strategies (none, equal, weighted, dynamic), sample counts (10, 20, 30, 40 per batch), and emoji modes (none, yes, weighted), as well as an error analysis with illustrative examples. The central claim is that the integrated pipeline yields a small but decisive improvement over prior systems while using about 3,336 training samples.

Significance. If the headline result holds, the paper would demonstrate a practical combination of MTL, active learning, and emoji-aware weighting for a low-resource Arabic NLP task, with thorough ablations and a useful qualitative error analysis. The strengths are the systematic grid of configurations, the explicit description of preprocessing, and the attention to emoji semantics. However, the load-bearing claims of state-of-the-art performance and sample efficiency are not yet supported by the evidence: the reported number is the best cell of a large grid with a single seed, the comparison table is incomplete, and the sample-counting methodology is inconsistent. These issues are central to the paper's contribution rather than presentation problems, so they must be addressed before the claims can be accepted.

major comments (3)
  1. [§4.1.1, Tables 4 and 7] The headline macro F1 of 85.42% is the maximum of a grid of 3 MTL weightings × 4 uncertainty strategies × 4 sample counts × 3 emoji modes (144 configurations), evaluated with a single random state (42). No variance, error bars, or significance tests are reported. Adjacent configurations are within 0.2–0.8 F1 points (e.g., Static/Equal at 85.19 and Dynamic/Weighted at 84.60 in Table 7), and the margin over the paper's own STL model (85.31, Table 5) is 0.11 points. These gaps are consistent with typical seed-level variance in transformer fine-tuning. The paper should report multiple seeds with means and standard deviations, and use a multiple-comparison-aware analysis to establish that the selected configuration is not a selection artifact.
  2. [§4.2, Tables 3 and 5] The claim of using "significantly fewer fine-tuning samples" is not measurable as presented. Table 3 defines "Total Fine-tuned Samples" as cumulative counts that include repeated instances across epochs, with values exceeding the dataset size (e.g., 17,114 > 8,557). Table 5 lists comparators' training sizes as unique or external corpus sizes (8,557, 19,906, 12,473), while the proposed MTL model's "3,336" is undefined and appears nowhere else in the paper. If 3,336 is a cumulative fine-tuning count, it is not comparable to unique training sizes; if it is unique samples, the active-learning loop that produces it is not specified in Section 3.2, which does not describe the number of selection iterations, how the per-batch count (10, 20, 30, 40) maps to the total, or how early stopping interacts with sample accumulation. The sample-efficiency claim is therefore unsupported.
  3. [§4.3, Table 5] The state-of-the-art claim is based on a selective comparison table that includes only five prior systems and omits the full OSACT2022 shared-task leaderboard. Several systems that participated in the shared task are not listed, so the reader cannot verify whether any system scored above 85.42%. Moreover, the margin over Mostafa et al. [35] is 0.25 F1 points, which is within the noise range identified in the first major comment. The paper should report the complete leaderboard and explicitly position the proposed model in that ranking, including the winning systems, before claiming state-of-the-art performance.
minor comments (6)
  1. [§4.2] The sentence "The highest F1 scores were obtained using the dynamic weighting approach" is imprecise; Table 4 shows that dynamic MTL weighting wins only with equal uncertainty (85.42), while with weighted uncertainty it reaches 84.60 and with none 84.01. Please state the configuration-specific nature of this conclusion.
  2. [§3.2 vs. §3.3] Equation (4) is introduced for dynamic entropy sampling, but Section 3.3 describes dynamic task weighting for the MTL loss as "based on loss" without providing an equation. The two dynamic mechanisms are easy to conflate; please explicitly distinguish dynamic loss weighting from dynamic uncertainty weighting and clarify how each is used.
  3. [Tables 6 and 7] The captions of Tables 6 and 7 use "MLT" instead of "MTL"; please correct this typo.
  4. [Throughout] There are several typographical errors: "ArabBERT" appears instead of "AraBERT" in the Section 4.2 text; "SoftMax" should be "softmax"; "Proceedinsg" appears in references [32]–[35]; "distinguishe" appears in Section 2. A careful proofread is recommended.
  5. [Table 3] The column "Total Fine-tuned Samples" is confusing because the selected sample counts (10, 20, 30, 40) do not obviously map to the reported totals (e.g., 5,360). Please clarify the active-learning schedule, including the number of epochs and how repeated instances are counted, either in the text or in the table caption.
  6. [§4.1.1] For reproducibility, please report all hyperparameters not currently specified, including the total number of training epochs, warm-up steps, and the exact procedure for initializing the uncertainty-based selection (e.g., whether the model is warm-started on a random subset).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework is evaluated empirically against external baselines, and the self-citations used are not load-bearing.

full rationale

The paper makes no derivation that reduces to its own inputs. The claimed contribution is an empirical pipeline combining entropy-based uncertainty sampling (Eqs. 1-5), MTL loss weighting (equal/static/dynamic), and weighted emoji handling, evaluated on a held-out OSACT2022 test set. The dynamic weight update in Eq. 4 uses the model's own dev macro-F1 to adjust training weights, but this is an optimization heuristic rather than a predicted quantity; the final reported 85.42% macro-F1 is an external test-set result, not a value forced by the weighting equation. The active-learning sample-count comparisons are training-budget descriptions, not fitted parameters renamed as predictions. The two self-citations ([49] and [54]) support the choice of a Twitter-pretrained BERT model and the general practice of weighting the main task more heavily; these are prior empirical claims rather than uniqueness theorems or definitional constraints, and the paper's state-of-the-art comparison rests on external baselines in Table 5. Weaknesses such as single-seed evaluation, best-of-grid selection, and incomplete leaderboard reporting are statistical and correctness concerns, not circularity. No specific equation or fitted input is shown to equal another by construction, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper's headline result depends on a number of empirically chosen hyperparameters and assumptions about the data labeling schema. The most consequential are the task weights, the dynamic weighting bounds, the undisclosed emoji list, and the assumption that vulgar and violent labels are independent of offensive. None of these are derived or independently validated, and the 'state-of-the-art' claim relies on a comparison table that may omit relevant systems.

free parameters (5)
  • Static task weights (offensive/vulgar/violent) = 0.7 / 0.15 / 0.15
    Empirically set in Section 3.3; the main result uses dynamic weighting but the static weighting ablations depend on these values.
  • Uncertainty sampling weights (W_off, W_vio, W_vul) = 2, 1, 1
    Set 'empirically' in Section 3.2; used in the weighted entropy acquisition function.
  • Dynamic weighting bounds and coefficients = T_min=0.75, W_min=0.5, W_max=2.0, coefficients 2/3 and 1/2
    Parameters in Eq. (4) and (5) chosen by hand; no sensitivity analysis is reported.
  • Weighted emoji list and weights = not disclosed
    A 'predefined list of emojis commonly associated with offensive speech' is mentioned but the list and weights are not given, making the mechanism unfalsifiable in the paper.
  • Number of selected samples in active learning = 10 (headline); grid up to 40
    The best result uses 10 selected samples per batch (Section 4.2); the choice is post hoc from Table 7.
assumptions (5)
  • domain assumption Auxiliary tasks (vulgar and violent) are sufficiently related to the main offensive task to improve it via hard sharing.
    Sections 3.3 and 4.2 assume MTL with these auxiliary tasks helps; no task relatedness analysis is provided.
  • domain assumption Vulgar and violent labels are independent of the offensive label.
    The paper excludes hate because it is only labeled for offensive tweets, but does not verify that vulgar and violent are independently labeled; low counts (132 and 60) suggest a possible hierarchy. Section 4.1.
  • domain assumption The OSACT2022 ground-truth labels are correct and the test set was not used to tune hyperparameters.
    Standard benchmark assumption; the paper gives no annotation-quality analysis.
  • domain assumption Entropy-based uncertainty sampling selects more informative samples than random sampling for this task.
    No random-sampling baseline is included; the active learning benefit is inferred from comparisons without a random-selection control.
  • domain assumption AraBERT-TwitterV2, pretrained on social media text, transfers well to OSACT2022 tweets.
    Used as the encoder in all experiments; no comparison to other Arabic PLMs is made in this work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-task Learning with Active Learning for Arabic Offensive Speech Detection." pith.science (2026). https://pith.science/paper/IR4EUWJ3

@misc{pith2026250602753,
  author       = {Pith},
  title        = {Pith review of: Multi-task Learning with Active Learning for Arabic Offensive Speech Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IR4EUWJ3}},
  note         = {Machine review of arXiv:2506.02753}
}
read the original abstract

The rapid growth of social media has amplified the spread of offensive, violent, and vulgar speech, which poses serious societal and cybersecurity concerns. Detecting such content in Arabic text is particularly complex due to limited labeled data, dialectal variations, and the language's inherent complexity. This paper proposes a novel framework that integrates multi-task learning (MTL) with active learning to enhance offensive speech detection in Arabic social media text. By jointly training on two auxiliary tasks, violent and vulgar speech, the model leverages shared representations to improve the detection accuracy of the offensive speech. Our approach dynamically adjusts task weights during training to balance the contribution of each task and optimize performance. To address the scarcity of labeled data, we employ an active learning strategy through several uncertainty sampling techniques to iteratively select the most informative samples for model training. We also introduce weighted emoji handling to better capture semantic cues. Experimental results on the OSACT2022 dataset show that the proposed framework achieves a state-of-the-art macro F1-score of 85.42%, outperforming existing methods while using significantly fewer fine-tuning samples. The findings of this study highlight the potential of integrating MTL with active learning for efficient and accurate offensive language detection in resource-constrained settings.

Figures

Figures reproduced from arXiv: 2506.02753 by the authors.

Figure 1
Figure 1. The Proposed Framework. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of the MTL component. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 47 canonical work pages

  1. [54]

    Enhancing stance detection through sequential weighted multi-task learning

    Nora Alturayeif, Hamzah Luqman, and Moataz Ahmed. Enhancing stance detection through sequential weighted multi-task learning. Social Network Analysis and Mining, 14(1):7, 2023. 14

    work page 2023

  2. [35]

    Gof at arabic hate speech 2022: breaking the loss function convention for data-imbalanced arabic offensive text detection

    Ali Mostafa, Omar Mohamed, and Ali Ashraf. Gof at arabic hate speech 2022: breaking the loss function convention for data-imbalanced arabic offensive text detection. In Proceedinsg of the 5th Workshop on Open- Source Arabic Corpora and Processing Tools with Shared Tasks on Qur’an QA and Fine-Grained Hate Speech Detection, pages 167–175, 2022

    work page 2022

  3. [1]

    Cyber-security: Identity deception detection on social media platforms

    Estee Van der Walt, Jan HP Eloff, and Jacomine Grobler. Cyber-security: Identity deception detection on social media platforms. Computers & Security, 78:76–89, 2018

    work page 2018

  4. [2]

    Automated hate speech detection and the problem of offensive language

    Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. Automated hate speech detection and the problem of offensive language. In Proceedings of the international AAAI conference on web and social media, volume 11, pages 512–515, 2017

    2017

  5. [3]

    A survey on automatic detection of hate speech in text

    Paula Fortuna and Sérgio Nunes. A survey on automatic detection of hate speech in text. ACM Computing Surveys (CSUR), 51(4):1–30, 2018

    work page 2018

  6. [4]

    Social media and the internet of things towards data-driven policymaking in the arab world: potential, limits and concerns

    Fadi Salem. Social media and the internet of things towards data-driven policymaking in the arab world: potential, limits and concerns. The Arab Social Media Report, Dubai: MBR School of Government, 7, 2017

    work page 2017

  7. [5]

    Wasm: A dataset for hashtag recommendation for arabic tweets

    Maged S Al-Shaibani, Hamzah Luqman, Abdulaziz S Al-Ghofaily, and Abdullatif A Al-Najim. Wasm: A dataset for hashtag recommendation for arabic tweets. Arabian Journal for Science and Engineering, 49(9):12131–12145, 2024

    work page 2024

  8. [6]

    An overview of multi-task learning

    Yu Zhang and Qiang Yang. An overview of multi-task learning. National Science Review, 5(1):30–43, 2018

    2018

Show all 54 references
  1. [7]

    Which tasks should be learned together in multi-task learning? In International conference on machine learning , pages 9120–9132

    Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Which tasks should be learned together in multi-task learning? In International conference on machine learning , pages 9120–9132. PMLR, 2020. 11 Running Title for Header

    2020

  2. [8]

    Empirical evaluation of multi-task learning in deep neural networks for natural language processing

    Jianquan Li, Xiaokang Liu, Wenpeng Yin, Min Yang, Liqun Ma, and Yaohong Jin. Empirical evaluation of multi-task learning in deep neural networks for natural language processing. Neural Computing and Applications, 33:4417–4428, 2021

    2021

  3. [9]

    and Yang Q

    Zhang Y . and Yang Q. A survey on multi-task learning.IEEE Transactions on Knowledge and Data Engineering, 34(12):5586–5609, 2018

    2018

  4. [10]

    Active learning for single-stage object detection in uav images

    Asma Yamani, Albandari Alyami, Hamzah Luqman, Bernard Ghanem, and Silvio Giancola. Active learning for single-stage object detection in uav images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1860–1869, 2024

    2024

  5. [11]

    Active learning for bert: an empirical study

    Liat Ein Dor, Alon Halfon, Ariel Gera, Eyal Shnarch, Lena Dankin, Leshem Choshen, Marina Danilevsky, Ranit Aharonov, Yoav Katz, and Noam Slonim. Active learning for bert: an empirical study. In Proceedings of the 2020 conference on empirical methods in natural language process...

    2020

  6. [12]

    Active learning literature survey

    Burr Settles. Active learning literature survey. 2009

    2009

  7. [13]

    Conditionally adaptive multi-task learning: Improving transfer learning in nlp using fewer parameters & less data

    Jonathan Pilault, Amine Elhattami, and Christopher Pal. Conditionally adaptive multi-task learning: Improving transfer learning in nlp using fewer parameters & less data. arXiv preprint arXiv:2009.09139, 2020

    2009 arXiv

  8. [14]

    Offensive language detection in arabic social networks using evolutionary-based classifiers learned from fine-tuned embeddings

    Fatima Shannaq, Bassam Hammo, Hossam Faris, and Pedro A Castillo-Valdivieso. Offensive language detection in arabic social networks using evolutionary-based classifiers learned from fine-tuned embeddings. IEEE Access, 10:75018–75039, 2022

    2022

  9. [15]

    Offensive language detection for low resource language using deep sequence model

    Anas Ali Khan, M Hammad Iqbal, Shibli Nisar, Awais Ahmad, and Waseem Iqbal. Offensive language detection for low resource language using deep sequence model. IEEE Transactions on Computational Social Systems, 2023

    2023

  10. [16]

    Detection of offensive language and its severity for low resource language

    Ramsha Saeed, Hammad Afzal, Sadaf Abdul Rauf, and Naima Iltaf. Detection of offensive language and its severity for low resource language. ACM Transactions on Asian and Low-Resource Language Information Processing, 22(6):1–27, 2023

    2023

  11. [17]

    Offensive language detection in nepali social media

    Nobal B Niraula, Saurab Dulal, and Diwa Koirala. Offensive language detection in nepali social media. In Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 67–75, 2021

    2021

  12. [18]

    Detect- ing hateful and offensive speech in arabic social media using transfer learning

    Zakaria Boulouard, Mariya Ouaissa, Mariyam Ouaissa, Moez Krichen, Mutiq Almutiq, and Karim Gasmi. Detect- ing hateful and offensive speech in arabic social media using transfer learning. Applied Sciences, 12(24):12823, 2022

    2022

  13. [19]

    Interpretable and high-performance hate and offensive speech detection

    Marzieh Babaeianjelodar, Gurram Poorna Prudhvi, Stephen Lorenz, Keyu Chen, Sumona Mondal, Soumyabrata Dey, and Navin Kumar. Interpretable and high-performance hate and offensive speech detection. In International Conference on Human-Computer Interaction, pages 233–244. Springer, 2022

    2022

  14. [20]

    Deep learning and natural language processing in computation for offensive language detection in online social networks by feature selection and ensemble classification techniques

    Manish Anand, Kishan Bhushan Sahay, Mohammed Altaf Ahmed, Daniyar Sultan, Radha Raman Chandan, and Bharat Singh. Deep learning and natural language processing in computation for offensive language detection in online social networks by feature selection and ensemble classifica...

    2023

  15. [21]

    Detecting offensive language based on graph attention networks and fusion features

    Zhenxiong Miao, Xingshu Chen, Haizhou Wang, Rui Tang, Zhou Yang, Tiemai Huang, and Wenyi Tang. Detecting offensive language based on graph attention networks and fusion features. IEEE Transactions on Computational Social Systems, 11(1):1493–1505, 2023

    2023

  16. [22]

    Osact4 shared task on offensive language detection: Intensive preprocessing-based approach

    Fatemah Husain. Osact4 shared task on offensive language detection: Intensive preprocessing-based approach. arXiv preprint arXiv:2005.07297, 2020

    2005 arXiv

  17. [23]

    Leveraging affective bidirectional transformers for offensive language detection

    AbdelRahim Elmadany, Chiyu Zhang, Muhammad Abdul-Mageed, and Azadeh Hashemi. Leveraging affective bidirectional transformers for offensive language detection. arXiv preprint arXiv:2006.01266, 2020

    2006 arXiv

  18. [24]

    Arabic offensive language detection with attention-based deep neural networks

    Bushr Haddad, Zoher Orabe, Anas Al-Abood, and Nada Ghneim. Arabic offensive language detection with attention-based deep neural networks. In Proceedings of the 4th workshop on open-source Arabic corpora and processing tools, with a shared task on offensive language detection, ...

    2020

  19. [25]

    Combining character and word embeddings for the detection of offensive language in arabic

    Abdullah I Alharbi and Mark Lee. Combining character and word embeddings for the detection of offensive language in arabic. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Offensive Language Detection, pages 91–96, 2020

    2020

  20. [26]

    Ks@ lth at semeval-2020 task 12: Fine-tuning multi-and monolingual transformer models for offensive language detection

    Kasper Socha. Ks@ lth at semeval-2020 task 12: Fine-tuning multi-and monolingual transformer models for offensive language detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 2045–2053, 2020

    2020

  21. [27]

    Kuisail at semeval-2020 task 12: Bert-cnn for offensive speech identification in social media

    Ali Safaya, Moutasem Abdullatif, and Deniz Yuret. Kuisail at semeval-2020 task 12: Bert-cnn for offensive speech identification in social media. arXiv preprint arXiv:2007.13184, 2020. 12 Running Title for Header

    2020 arXiv

  22. [28]

    Salamnet at semeval-2020 task12: Deep learning approach for arabic offensive language detection

    Fatemah Husain, Jooyeon Lee, Samuel Henry, and Ozlem Uzuner. Salamnet at semeval-2020 task12: Deep learning approach for arabic offensive language detection. arXiv preprint arXiv:2007.13974, 2020

    2020 arXiv

  23. [29]

    Semeval-2020 task 12: Multilingual offensive language identification in social media (offenseval 2020)

    Marcos Zampieri, Preslav Nakov, Sara Rosenthal, Pepa Atanasova, Georgi Karadzhov, Hamdy Mubarak, Leon Der- czynski, Zeses Pitenis, and Ça˘grı Çöltekin. Semeval-2020 task 12: Multilingual offensive language identification in social media (offenseval 2020). arXiv preprint arXiv:...

    2020 arXiv

  24. [30]

    Alt at semeval-2020 task 12: Arabic and english offensive language identification in social media

    Sabit Hassan, Younes Samih, Hamdy Mubarak, and Ahmed Abdelali. Alt at semeval-2020 task 12: Arabic and english offensive language identification in social media. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 1891–1897, 2020

    2020

  25. [31]

    Bhamnlp at semeval-2020 task 12: An ensemble of different word embeddings and emotion transfer learning for arabic offensive language identification in social media

    Abdullah I Alharbi and Mark Lee. Bhamnlp at semeval-2020 task 12: An ensemble of different word embeddings and emotion transfer learning for arabic offensive language identification in social media. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 1532–1...

    2020

  26. [32]

    Upv at the arabic hate speech 2022 shared task: Offensive language and hate speech detection using transformers and ensemble models

    Angel Felipe Magnossao de Paula, Paolo Rosso, Imene Bensalem, and Wajdi Zaghouani. Upv at the arabic hate speech 2022 shared task: Offensive language and hate speech detection using transformers and ensemble models. In Proceedinsg of the 5th Workshop on Open-Source Arabic Corp...

    2022

  27. [33]

    Overview of osact5 shared task on arabic offensive language and hate speech detection

    Hamdy Mubarak, Hend Al-Khalifa, and AbdulMohsen Al-Thubaity. Overview of osact5 shared task on arabic offensive language and hate speech detection. In Proceedinsg of the 5th Workshop on Open-Source Arabic Corpora and Processing Tools with Shared Tasks on Qur’an QA and Fine-Gra...

    2022

  28. [34]

    Guct at arabic hate speech 2022: Towards a better isotropy for hatespeech detection

    Nehal Elkaref and Mervat Abu-Elkheir. Guct at arabic hate speech 2022: Towards a better isotropy for hatespeech detection. In Proceedinsg of the 5th Workshop on Open-Source Arabic Corpora and Processing Tools with Shared Tasks on Qur’an QA and Fine-Grained Hate Speech Detectio...

    2022

  29. [36]

    Detection of arabic offensive language in social media using machine learning models

    Aya Mousa, Ismail Shahin, Ali Bou Nassif, and Ashraf Elnagar. Detection of arabic offensive language in social media using machine learning models. Intelligent Systems with Applications, 22:200376, 2024

    2024

  30. [37]

    Latent multi-task architecture learning

    Sebastian Ruder, Joachim Bingel, Isabelle Augenstein, and Anders Søgaard. Latent multi-task architecture learning. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 4822–4829, 2019

    2019

  31. [38]

    Multi-task learning with deep neural networks: A survey

    Michael Crawshaw. Multi-task learning with deep neural networks: A survey. arXiv preprint arXiv:2009.09796, 2020

    2009 arXiv

  32. [39]

    Exploring multi-task multi-lingual learning of transformer models for hate speech and offensive speech identification in social media

    Sudhanshu Mishra, Shivangi Prasad, and Shubhanshu Mishra. Exploring multi-task multi-lingual learning of transformer models for hate speech and offensive speech identification in social media. SN Computer Science, 2:1–19, 2021

    2021

  33. [40]

    Overview of the hasoc track at fire 2019: Hate speech and offensive content identification in indo-european languages

    Thomas Mandl, Sandip Modha, Prasenjit Majumder, Daksh Patel, Mohana Dave, Chintak Mandlia, and Aditya Patel. Overview of the hasoc track at fire 2019: Hate speech and offensive content identification in indo-european languages. In Proceedings of the 11th annual meeting of the ...

    2019

  34. [41]

    Multi-task learning in under-resourced dravidian languages

    Adeep Hande, Siddhanth U Hegde, and Bharathi Raja Chakravarthi. Multi-task learning in under-resourced dravidian languages. Journal of Data, Information and Management, 4(2):137–165, 2022

    2022

  35. [42]

    Integrating implicit and explicit linguistic phenomena via multi-task learning for offensive language detection

    Flor Miriam Plaza-del Arco, M Dolores Molina-González, L Alfonso Ureña-López, and María-Teresa Martín- Valdivia. Integrating implicit and explicit linguistic phenomena via multi-task learning for offensive language detection. Knowledge-Based Systems, 258:109965, 2022

    2022

  36. [43]

    Offensive language identification with multi-task learning

    Marcos Zampieri, Tharindu Ranasinghe, Diptanu Sarkar, and Alex Ororbia. Offensive language identification with multi-task learning. Journal of Intelligent Information Systems, 60(3):613–630, 2023

    2023

  37. [44]

    Multi-task learning using arabert for offensive language detection

    Marc Djandji, Fady Baly, Wissam Antoun, and Hazem Hajj. Multi-task learning using arabert for offensive language detection. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Offensive Language Detection, pages 97–101, 2020

    2020

  38. [45]

    icompass at arabic hate speech 2022: Detect hate speech using qrnn and transformers

    Mohamed Aziz Bennessir, Malek Rhouma, Hatem Haddad, and Chayma Fourati. icompass at arabic hate speech 2022: Detect hate speech using qrnn and transformers. In Proceedinsg of the 5th Workshop on Open-Source Arabic Corpora and Processing Tools with Shared Tasks on Qur’an QA and...

    2022

  39. [46]

    Alexu-aic at arabic hate speech 2022: Contrast to classify

    Ahmad Shapiro, Ayman Khalafallah, and Marwan Torki. Alexu-aic at arabic hate speech 2022: Contrast to classify. arXiv preprint arXiv:2207.08557, 2022

    2022 arXiv

  40. [47]

    Emojis as anchors to detect arabic offensive language and hate speech

    Hamdy Mubarak, Sabit Hassan, and Shammur Absar Chowdhury. Emojis as anchors to detect arabic offensive language and hate speech. Natural Language Engineering, 29(6):1436–1457, 2023

    2023

  41. [48]

    Arabert: Transformer-based model for arabic language understand- ing

    Wissam Antoun, Fady Baly, and Hazem Hajj. Arabert: Transformer-based model for arabic language understand- ing. arXiv preprint arXiv:2003.00104, 2020

    2003 arXiv

  42. [49]

    Fine-grained sentiment analysis of arabic covid-19 tweets using bert-based transformers and dynamically weighted loss function

    Nora Alturayeif and Hamzah Luqman. Fine-grained sentiment analysis of arabic covid-19 tweets using bert-based transformers and dynamically weighted loss function. Applied Sciences, 11(22):10694, 2021

    2021

  43. [50]

    Deep bayesian active learning for natural language processing: Results of a large-scale empirical study

    Aditya Siddhant and Zachary C Lipton. Deep bayesian active learning for natural language processing: Results of a large-scale empirical study. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2904–2909, 2018

    2018

  44. [51]

    Active learning using adaptable task-based prioritisation

    Shaheer U Saeed, João Ramalhinho, Mark Pinnock, Ziyi Shen, Yunguan Fu, Nina Montaña-Brown, Ester Bonmati, Dean C Barratt, Stephen P Pereira, Brian Davidson, et al. Active learning using adaptable task-based prioritisation. Medical Image Analysis, 95:103181, 2024

    2024

  45. [52]

    Active learning and crowdsourcing for machine translation in low resource scenarios

    Vamshi Ambati. Active learning and crowdsourcing for machine translation in low resource scenarios. Carnegie Mellon University, 2012

    2012

  46. [53]

    How to measure uncertainty in uncertainty sampling for active learning

    Vu-Linh Nguyen, Mohammad Hossein Shaker, and Eyke Hüllermeier. How to measure uncertainty in uncertainty sampling for active learning. Machine Learning, 111(1):89–122, 2022

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.