An AI system spots offensive Arabic speech with fewer labeled examples

A model trained on offensive, violent and vulgar language reached a macro F1-score of 85.42% on the OSACT2022 dataset.

Editorial illustration — not from the study.

The system was trained to detect offensive, violent and vulgar speech together, allowing the tasks to share information. It also used active learning: uncertainty-based methods selected samples considered most useful for further training, rather than relying on large numbers of labeled examples from the start. Emojis linked to offensive speech were given additional weight.

On the OSACT2022 dataset, the framework achieved a reported macro F1-score of 85.42% and outperformed the comparison methods described by the researchers. The strongest results came from dynamically balancing the related tasks and using average uncertainty sampling, according to the study.

What the system found

The researchers report a macro F1-score of 85.42% for their framework on the OSACT2022 dataset, calling it a state-of-the-art result in their comparison. The multi-task system performed better than a single-task version when only a small number of selected samples were used for fine-tuning. Dynamic task weighting produced higher scores than equal weighting, while average uncertainty sampling combined with dynamic weighting produced the best result among the tested strategies. Giving greater weight to selected emojis also improved performance compared with leaving emojis out; one reported configuration reached an F1-score of 84.84% with 20 samples per batch in each training epoch.

Why fewer examples matter

Arabic offensive-speech detection is difficult when labeled examples are scarce, partly because Arabic social-media writing varies across dialects and often depends on context, sarcasm and emojis. A system that can learn effectively from fewer selected examples could make this type of language technology more practical in settings with limited labeled data. The study suggests that training on related forms of harmful language can help the main detection task, although it does not establish how the system would perform beyond the tested dataset.

Evidence and limits

The evidence comes from experiments on the OSACT2022 dataset, evaluated with macro F1-score, a measure that summarizes performance across classes. The researchers compared single-task and multi-task setups, several task-weighting approaches and uncertainty-based sample-selection strategies. The reported results are therefore experimental rather than a test of the system in live social-media moderation.

The model still had difficulty with implicit sarcasm, cultural references and context-dependent language. The authors also note possible labeling inconsistencies. Generalizability remains untested on other Arabic offensive-language datasets or multilingual benchmarks, so the reported score should not be treated as a measure of performance across all Arabic online speech.

// Source

ACM Transactions on Asian and Low-Resource Language Information Processing · 2026 · DOI: 10.1145/3848627

View paper (DOI)Open access versionPublished 2026-09-22

Authors: Aisha Alansari, Hamzah Luqman

Institutions: King Fahd University of Petroleum and Minerals

Programmable hardware speeds matrix calculations up to 6.3 times

Researchers tested a programmable hardware design that imitates selected quantum-computing ideas to run matrix calculations in parallel. In representative benchmarks, the design reached up to 6.3 times the speed and 4.1 times the energy efficiency of conventional matrix co-processors and embedded graphics processors.

2 days ago·13 views

AI & Computing

AI and docking agreed in some drug screens, but not all

Researchers evaluated three AI systems on hundreds of molecules bound to a SARS-CoV-2 protein and in three prospective drug screens. The results suggest that combining AI-based predictions with traditional docking may help prioritize candidates, although no single approach worked best in every test.

4 days ago·50 views

AI & Computing

A hardware monitor detected simulated attacks on car networks in 14 microseconds

Researchers developed a hardware-assisted system intended to detect message injection, spoofing, flooding and impersonation attacks on automotive networks. In tests using simulated traffic and a prototype automotive processor system, it reported an average detection time of 14 microseconds and less than 2 percent CPU overhead.

5 days ago·43 views