Translate Vertical Text and Mixed Scripts

Identify the reading directions

Look for columns that run from top to bottom, side notes that read horizontally, and captions placed beside an image. A single page may use more than one direction. OCR can recognize every character but still return the lines in the wrong sequence.

Keep the full page as a reference, then make separate crops for text regions with different directions. Include a small overlap so you can reconstruct their positions.

Separate scripts by function

A label may combine a local name, a Latin transliteration, a model number, and an English slogan. Translate only the parts that carry meaning in context. Product codes, personal names, and established brand names often need to remain unchanged.

Compare punctuation and brackets carefully because they can indicate pronunciation, annotation, or a separate translation rather than part of the main sentence.

Dense character layouts need additional capture choices covered in the guide to translating Chinese, Japanese, and Korean text in images.

Rebuild sequence before meaning

Arrange the extracted regions in the order a reader would encounter them, then translate. Preserve headings and captions as separate blocks. This prevents a side note from being inserted into the middle of a sentence.

If the order remains uncertain, provide the crops to a reader familiar with the script. Translation should follow the document’s structure, not attempt to hide uncertainty with smooth prose.