Reproduce “EDRAC: Benchmarking Arabic Dialect Reading Comprehension” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Noor Abo Mokh, Kirill Chirkunov, Teresa Lynn, Nizar Habash, Reham Marzouk, Malik H. Altakrori, Younes Samih, Muhammed Abu Odeh and 10 more
Researchers developed EDRAC, the first large-scale benchmark for generative Machine Reading Comprehension in five Arabic dialects, built from natural spoken interactions using a human-LLM pipeline. Benchmarking revealed that while top models achieve high semantic accuracy, they struggle to produce dialectally natural outputs, highlighting limitations in current evaluation metrics for dialectal fidelity.
Key insights
- A human-LLM collaborative pipeline can efficiently produce high-quality, culturally grounded datasets for under-resourced language varieties, effectively mitigating the challenges of manual annotation.
- Current LLMs, even high-performing ones, exhibit a significant gap between generating semantically accurate answers and producing outputs that are dialectally natural and authentic to native speakers.
- Standard automatic evaluation metrics (ROUGE-L, BERTScore) are insufficient for adequately assessing dialectal fidelity in generative tasks, often over-rewarding semantic similarity while overlooking crucial nuances of dialectal expression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.