AraSEG corpus benchmarks Arabic sentence segmentation across 8 genres

AraSEG is a genre-diverse Arabic sentence segmentation corpus covering eight genres and varied punctuation conditions, revealing that lightweight encoders and dependency parser models outperform LLMs in the hardest low-punctuation settings. Cross-genre generalization remains a persistent challenge even with increased training data.