Skip to content
conftrace_
2026 EMNLP EMNLP 2026

From Subwords to Pixels: Evaluating Text Representations for Arabic Machine Translation

AbstractSubword tokenization underpins modern neural NLP but can be suboptimal for morphologically rich languages with substantial orthographic variation. In this paper, we compare vocabulary-based and vocabulary-free representations for Arabic-to-English machine translation, including subword, byte-level, and pixel-level approaches. Across 35 evaluation sets spanning Modern Standard Arabic (MSA) and 18 dialects, all approaches achieve comparable translation performance with sufficient supervision. While subword models perform best with limited MSA pretraining data, differences diminish with scale and dialect-specific fine-tuning. Pixel-based models further exhibit the greatest robustness to unseen orthographic variation and provide the highest training and inference throughput. We make our code and models publicly available.

More from EMNLP 2026

All EMNLP 2026 papers →