Wajdi Zaghouani

Transformer models excel in detecting polarized stances in social media, but significant challenges remain in generalizing across topics and predicting neutral positions.

LLMs could either preserve or erase dialects, and this paper offers 12 guidelines to ensure they empower rather than homogenize linguistic diversity.

LLM safety evaluations in English miss critical vulnerabilities: GPT-4o's refusal rate on Kazakh safety prompts varies from 5.5% to 53.8% across categories.

LLMs can supercharge democratic deliberation, but only if humans retain ultimate control through a carefully scaffolded, accountable system that prevents AI from steamrolling diverse perspectives.

LLM safety evaluations now have a crucial resource for Albanian, a significantly underrepresented low-resource language.

A new corpus of Arabic social media data offers a unique lens into audience sentiment and engagement around women's empowerment across diverse Arabic dialects and countries.

Conflict wins: Arabic social media users engage with divisive posts 2-4x more than unifying ones, according to a new dataset of 6,000 Facebook posts.

Finally, a large-scale, ecologically valid Arabic dataset lets researchers study the interplay of discriminatory language and audience response on Facebook.