Work

Table 7. Three-Seed Transformer Official Test Results. - https://doi.org/10.70594/brain/17.3/3

Year: 2026

Type: supplementary-materials

Abstract: Across individual seeds, Urdu-RoBERTa had the highest mean Macro-F1 (0.8747 ± 0.0096)and the lowest mean fitting time among the three transformer families. XLM-R had the highestmean ROC-AUC (0.9480 ± 0.0019) and Hate-class precision (0.9047 ± 0.0053), but lower recall.The clean mBERT rerun obtained a mean Macro-F1 of 0.8640 ± 0.0076 and a mean ROC-AUC of0.9400 ± 0.0020. Probability aggregation improved family-level performance, with Urdu-RoBERTareaching Macro-F1 = 0.8836; XLM-R remained strongest by ROC-AUC (0.9529), and cleanmBERT reached aggregate Macro-F1 = 0.8722.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.23795982

Authors: Ullah, Haseeb, Saleem, Husnain, Kanwal, Asia, Javed, Muhammad, Khan, Hamid Masood, Hanif, Kiran

Institutions: Gomal University, The University Of Agriculture Dera Ismail Khan

Language: English

Topic: Hate Speech and Cyberbullying Detection

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7224836588

Cites:

Cited by:

Related to: 10