AraHopeCorpus: Annotation Guidelines and Dataset for Hope Speech in Arabic Social Media Crisis Discourse

The dataset's significance isn't just scale or language coverage. The 64% hopeful classification suggests that crisis discourse on Arabic social media may be structurally different from English-language equivalents, where negativity bias often dominates. This raises a methodological question: are models trained on English hate speech datasets systematically blind to non-Western communication patterns that prioritize collective resilience over individual grievance?

This work sits directly alongside the Cultural Adaptation in Large Language Models paper from this week, which identified systematic failures when English-trained systems encounter non-Western discourse norms. AraHopeCorpus operationalizes that insight by building a dataset that reflects actual Arabic communication semantics rather than translating English toxicity frameworks. The ClimateChat dataset released the same day also demonstrates how large-scale social media corpora now routinely capture emotional framing and semantic themes, but AraHopeCorpus adds a critical dimension: it treats constructive language as a first-class annotation target rather than a residual category.

If teams building content moderation systems for Arabic-language platforms adopt this corpus and report measurable improvements in identifying constructive speech without increasing false negatives on actual harm, that validates the premise. Conversely, if the dataset remains primarily academic and moderation systems continue using English-derived toxicity models, it suggests the infrastructure gap between research and deployment remains wider than the dataset alone can bridge.