https://jsl5710.github.io/authors/uchendu-uchendu/Uchendu UchenduHugoBlox Kit (https://hugoblox.com)en-usWed, 08 Apr 2026 00:00:00 +0000https://jsl5710.github.io/publication/conference-paper-dia-harm/Wed, 08 Apr 2026 00:00:00 +0000https://jsl5710.github.io/publication/conference-paper-dia-harm/ <blockquote class="border-l-4 border-neutral-300 dark:border-neutral-600 pl-4 italic text-neutral-600 dark:text-neutral-400 my-6"> <p>🏆 <strong>Social Impact Paper Award — ACL 2026</strong>, San Diego, July 2–7, 2026.</p> </blockquote> <p> <figure > <div class="flex justify-center "> <div class="w-full" > <img alt="Award ceremony at ACL 2026" srcset="https://jsl5710.github.io/publication/conference-paper-dia-harm/award_hu_80ed752c554010bb.webp 320w, https://jsl5710.github.io/publication/conference-paper-dia-harm/award_hu_60d9b5863648e4e5.webp 480w, https://jsl5710.github.io/publication/conference-paper-dia-harm/award_hu_ed1787f737c5acbb.webp 760w" sizes="(max-width: 480px) 100vw, (max-width: 768px) 90vw, (max-width: 1024px) 80vw, 760px" src="https://jsl5710.github.io/publication/conference-paper-dia-harm/award_hu_80ed752c554010bb.webp" width="760" height="570" loading="lazy" data-zoomable /></div> </div></figure> <figure > <div class="flex justify-center "> <div class="w-full" > <img alt="Award recipients slide showing DIA-HARM" srcset="https://jsl5710.github.io/publication/conference-paper-dia-harm/award-slide_hu_6c685cfd29050812.webp 320w, https://jsl5710.github.io/publication/conference-paper-dia-harm/award-slide_hu_a0e9c278522b6ee6.webp 480w, https://jsl5710.github.io/publication/conference-paper-dia-harm/award-slide_hu_906922782ba7a4e0.webp 760w" sizes="(max-width: 480px) 100vw, (max-width: 768px) 90vw, (max-width: 1024px) 80vw, 760px" src="https://jsl5710.github.io/publication/conference-paper-dia-harm/award-slide_hu_6c685cfd29050812.webp" width="760" height="570" loading="lazy" data-zoomable /></div> </div></figure> </p> <p><strong>DIA-HARM</strong> investigates the robustness of harmful content detection systems across <strong>50 English dialects</strong>, addressing critical equity gaps in automated content moderation.</p> <p>Key contributions include:</p> <ul> <li><strong>Dialect-Diverse Detection (D3) Corpus</strong>: Over 195K samples derived from benchmark harmful content datasets, transformed using 189 morphosyntactic rules from eWAVE covering 50 dialects</li> <li><strong>Comprehensive Model Evaluation</strong>: 16 detection models tested — 10 fine-tuned, 5 zero-shot, and 1 in-context learning — revealing systematic performance disparities across dialect groups</li> <li><strong>D-PURIFY Validation</strong>: Quality filtering pipeline ensuring linguistic validity of dialect transformations with 97.2% average F1 for the best model (mDeBERTa)</li> <li><strong>Key Finding</strong>: Detection degradation correlates with density of morphosyntactic transformations rather than specific dialect features, with 2,450 dialect pairs analyzed</li> </ul> <p>Resources:</p> <ul> <li> </li> <li> </li> <li> </li> </ul>