EUVIMEDEUVIMED entdecken
Crossref

A Hybrid Deberta–Charcnn Multi-Task Framework for Robust Hate Speech Detection

Jayshree Kalawa, Arpana Chourasia

International Journal of Drug Delivery Technology · 21.05.2026 · 16 (29s)

DOI: 10.25258/ijddt.16.29s.77 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Drug Delivery Technology, online erschienen am 21.05.2026. Verfasst von Jayshree Kalawa und 1 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Growing numbers of online communication tools have facilitated hate speech at an unprecedented level leading to grave threats for digital safety and social cohesion. Hate speech detection is still challenging given implicit expressions, sarcasm, culture-specific vocabulary extension or words and artificial trends like coded slurs. While traditional rule-based and classical machine learning systems tend to be based on shallow lexical features the majority of these are still lacking in many transformer- based single-task models that either struggle with disguised hate or context-dependent hateful content. To tackle these challenges, we design a Hybrid-2 deep learning model that integrates contextual semantic modeling and character-level pattern learning with auxiliary supervision. The proposed model incorporates a DeBERTa-v3-base transformer encoder for learning deep contextual representation, a Character-CNN module to learn coded and misspelled hate patterns, and multi-task learning with pseudo sentiment and emotion label, respectively to enhance feature learning without auxiliary human labels. Adaptive SoftMax weighting is used to dynamically weight the primary hate classification task and auxiliary tasks in order to promote robustness and stability during training. The model has been tested with a large scale HateSpeechDatasetBalanced dataset from Kaggle (726.119 labeled text samples in balanced hate and none-hate) distribution for both. Experimental evaluations show the proposed approach can achieve a competitive performance of 93.78% for accuracy,94.04% for precision, 93.53% recall and an F1-score of 93.79%, which indicates robust discrimination between hateful and non-hateful content. The highest Performance-and-m-ebc AUC in our proposed model (89.5%) is 6 to 9 points% better than that observed in classical ML baselines such as SVM (≈84–88%) and improves robustness to obfuscation and manipulation of emotional contents.

Zitieren

APA

Kalawa, J., & Chourasia, A. (2026). A Hybrid Deberta–Charcnn Multi-Task Framework for Robust Hate Speech Detection. International Journal of Drug Delivery Technology, 16(29s). https://doi.org/10.25258/ijddt.16.29s.77

Vancouver

Kalawa J, Chourasia A. A Hybrid Deberta–Charcnn Multi-Task Framework for Robust Hate Speech Detection. International Journal of Drug Delivery Technology. 2026;16(29s). doi:10.25258/ijddt.16.29s.77

BibTeX
@article{Kalawa2026,
  author = {Kalawa, Jayshree and Chourasia, Arpana},
  title = {A Hybrid Deberta–Charcnn Multi-Task Framework for Robust Hate Speech Detection},
  journal = {International Journal of Drug Delivery Technology},
  year = {2026},
  volume = {16},
  number = {29s},
  doi = {10.25258/ijddt.16.29s.77}
}
Kostenlos starten →