Helsinki-NLP

/

opus-mt-ar-en

like 49

Follow

Helsinki-NLP Research Group 991

Ran a quick benchmark on OPUS-MT Arabic-English with 9 test cases across formal, technical, and dialectal inputs.

Key findings:

• MSA (Modern Standard Arabic): Strong performance, 3–14s latency
• Technical content: Handles ML/API terminology well, preserves code-switching
• Dialectal Arabic: Significant truncation — Egyptian "إزيك؟ كله تمام؟" reduced to "I was gonna ask you something" (missed the greeting entirely)
• Sudanese "يا زول" outputted as "Hey, Zol" with untranslated term

Latency range: 0.4s (simple) to 13.5s (technical sentences)

For production Arabic NLP pipelines, OPUS-MT works well for MSA but dialectal preprocessing would improve coverage.

Has anyone tested this against NLLB-200 for Arabic dialect coverage?