Vertical Japanese Text Translation: Layout and OCR Challenges¶

A page of vertical Japanese can look perfectly readable until OCR turns its columns into horizontal strings and puts the first column at the end. The translation may contain the right characters but still tell the reader to follow the wrong path through the page.

Japanese vertical writing, commonly called tategaki, isn’t simply horizontal writing turned on its side. The writing direction affects how readers move through columns, how punctuation and Latin text sit on the page, and how annotations relate to the words they explain. Translation therefore involves more than converting Japanese sentences into English. The document’s structure has to survive too.

This guide explains how vertical Japanese text works, why OCR and layout extraction can fail, and what to check when converting a tategaki document for readers who expect horizontal English. It also covers mixed-direction pages, ruby and furigana, line-breaking constraints, and the situations where preserving the source layout matters less than making a clear, accessible target document.

What vertical Japanese writing means¶

Japanese composition supports both horizontal and vertical writing modes. In the principal vertical mode described in Japanese layout guidance, characters run from top to bottom, columns progress from right to left, and readers move through the page in a different sequence from a conventional Western horizontal book. The W3C Requirements for Japanese Text Layout describes Japanese composition requirements for technologies such as CSS, SVG, XSL-FO, and ebook standards.

The basic distinction is easy to miss: tategaki isn’t a row of ordinary horizontal Japanese text rotated as one block. The characters have individual positions within vertical lines, and the page has a specific column sequence. A translation workflow that treats every visible mark as a single stream can preserve the characters while losing the organization that makes them readable.

The contrast is visible in a typical two-column example. A reader starts at the top of the right-hand column, follows it down, then moves to the top of the column to its left. A horizontal English reader usually expects a line to run from left to right, followed by the line below. If software extracts the Japanese by grouping text according to horizontal proximity, the resulting sequence may not follow either the Japanese page or the intended translated reading order.

Japanese text also uses square character frames for ideographs, hiragana, and katakana in the composition principle described by JLReq. That doesn’t mean every character occupies an identical visible shape or that every layout is a rigid grid. It does mean that spacing and alignment aren’t equivalent to inserting spaces between words in English. Removing or adding spaces indiscriminately can change the appearance of a Japanese line.

JLReq is a guide to composition requirements, not an automatic OCR recipe. The W3C document says its aim is to describe important composition issues and requirements, rather than propose implementation solutions. A translator or document team can use it to understand the rules a layout should respect, but shouldn’t expect it to specify a ready-made process for extracting every PDF or scan.

A Japanese EPUB working draft offers a useful description of the principal reading direction: “In vertical writing mode, reading systems should arrange characters from top to bottom, and arrange lines from right to left.” The wording appears in the 2010 first working draft on Japanese EPUB layout, so it describes a convention rather than a current conformance rule. For translation work, the practical point is that a page’s reading order must be identified before its text is rearranged.

A source document also can mix directions on the same page. The draft notes that running heads, page numbers, captions, and table entries commonly remain horizontal in a vertically set book. Boxed paragraphs may also use horizontal writing even when the main text runs vertically. A page-level label such as “vertical” therefore doesn’t tell an OCR system how to read every region.

That mix matters when the target language uses a different page logic. Translating the main narrative into left-to-right paragraphs may be sensible, while a caption or table still needs to be matched to the correct illustration or entry. If extraction flattens the page, the translated text can end up associated with the wrong object even when every sentence is accurate.

How vertical Japanese translation works in practice¶

A reliable process starts with the page, not the text box. Before translation, identify the document type, its dominant writing direction, and the regions that follow another direction. Then separate text recognition, translation, and layout review so that an error in one stage doesn’t silently become an error in the next.

1. Identify the document and its reading order¶

Begin by checking whether the source is editable text or an image of a page. An editable file may retain paragraphs, cells, and text direction as structured information. A scan or photograph contains pixels instead, so software must first infer where the characters and text regions are.

Next, inspect how the main content flows. For vertical Japanese, mark the order of columns from right to left and the direction within each column from top to bottom. Note any headings, captions, page numbers, tables, or boxed passages that read horizontally. Don’t assume a page’s largest text block represents every region on that page.

A two-column page illustrates the risk. If the first visible block in a plain-text extraction belongs to the left column, the extracted sequence may begin in the wrong place. If a caption sits beside an image, its proximity to the main text doesn’t make it part of the same paragraph. A layout map, even a simple one, gives a reviewer something concrete to compare with the recognition output.

2. Check recognition before translating¶

OCR means optical character recognition: software reads text from an image and produces machine-readable characters. Recognition of vertical Japanese requires more than identifying glyphs. The software also has to locate columns, decide which characters belong together, and establish the order in which those groups should be read.

A community discussion about vertical Japanese in EasyOCR describes a practical failure mode. One contributor says a detector that groups characters by horizontal proximity can interpret vertical columns as horizontal lines. Another commenter reports that rotating the image and then detecting text didn’t fully solve the problem. These reports concern a particular software discussion, not a benchmark of all OCR systems, but they show why image rotation alone isn’t a complete fix.

One commenter puts the grouping problem this way:

“The point is, craft by default try to bind characters that is horizontally near into lines, so when you write like c b a / c b a / c b a craft will understand them as 3 lines of cba”

The example describes a system that sees horizontal proximity rather than the page’s vertical columns. For the translator, the lesson is to inspect the extracted text and its reading order before treating it as a trustworthy source. A page that looks rotated correctly can still have its characters grouped into the wrong lines.

A 2026 LREC paper evaluated multimodal language models on synthetic images of horizontal and vertical Japanese, as well as real-world images of vertical Japanese documents. The paper reports that the evaluated models performed worse on vertical than horizontal Japanese in those datasets (Sasagawa, Kurita, and Kawahara, May 2026). That finding applies to the models and evaluation described in the study, not automatically to every OCR product or workflow.

The same paper reports that training with its synthetic Japanese OCR dataset improved performance for models that previously couldn’t handle vertical writing. That result is also specific to the study. For a document team, it reinforces the need to test the actual recognition setup on representative pages rather than infer performance from a tool’s general ability to read Japanese.

3. Preserve mixed-direction regions¶

After recognition, compare the extracted text with the original page by region. Keep the main vertical text separate from horizontal running heads, page numbers, captions, table entries, and boxed text. Connect each region to its place on the page, not just to a position in a long text string.

A table is a good example. A Japanese table entry may be horizontal even when the surrounding prose runs vertically. Extracting every character in one sequence can detach a cell value from its row or column label. A reviewer should verify the relationship between labels and entries before translating and again after the translated layout is built.

Illustrations create a similar problem. A caption near an image belongs with that image, even if a layout parser finds other text nearer to it in reading order. JLReq includes page design topics such as heading placement, illustrations, tables, notes, and reference marks. Those elements matter because a document communicates through its composition as well as through the character sequence.

Teams working with several text regions can benefit from a page-by-page review routine. Record which direction each region uses, identify its reading order, and check how it relates to surrounding objects. The process doesn’t need to force every source element into the same target layout. It does need to keep the meaning and associations clear.

4. Translate with layout decisions in view¶

A translation can be accurate as prose and still fail as a document. English often takes more space than Japanese for the same message, so a translated paragraph may not fit the source column or text box. The right response depends on the purpose of the deliverable: preserve a close visual match, or prioritize a clear and accessible target-language document.

Before choosing, decide what the reader needs to do with the output. A reference edition, a form, and a text prepared for general reading may call for different layout choices. A reference edition may need to keep each translation passage aligned with the corresponding source region. A general reading copy may work better as ordinary horizontal paragraphs, provided that headings, figures, notes, and references remain connected to the right content.

Keep text roles visible while translation is underway. A caption shouldn’t become an unlabelled paragraph; a page number shouldn’t join a sentence; a table entry shouldn’t drift away from its heading. These connections are easier to protect when translators and layout reviewers can see the source page and its regions, rather than working from a flattened text dump.

A document team can use a bilingual review sheet that places each source block beside its translation and identifies the block’s location. For example, the reviewer might mark “rightmost main column,” “horizontal caption under illustration,” or “table header.” Those labels make it possible to notice that the translated caption was attached to the wrong figure or that a column has been reviewed in the wrong order.

5. Rebuild and compare the target layout¶

Once the translation is ready, check the target document against the source and the translation itself. Verify that no source region disappeared during extraction, that the translated text belongs to the right region, and that tables, illustrations, notes, and references still relate to the right content.

A page-by-page comparison catches issues that a spell-checker won’t. A spell-checker can flag a misspelled word, but it won’t tell you that a caption sits below the wrong illustration or that a note refers to a different paragraph after reflow. A reviewer who knows both the language and the page structure can check content and positioning together.

When the translated document doesn’t need to reproduce tategaki, don’t preserve vertical columns just to imitate the source. Reformatting into horizontal English may improve readability. The requirement is to carry over the document’s information and relationships, not to create a visual facsimile at any cost.

For more on checking a translated file page by page, see translated PDF layout verification steps. A separate guide to preserving Word styles when translating can help teams keep headings and document structure intact after translation.

When preserving tategaki matters¶

Keeping the vertical layout can be useful when readers need to compare source and translation in the same visual structure. Examples include reference editions, language-learning material, annotated documents, and publications where page composition is part of the reading experience. In those cases, layout carries information that a plain text version may not preserve.

A bilingual edition may need a reader to find a particular source passage and its translation quickly. If the translated page follows a different order without clear labels or anchors, the reader may struggle to compare the two versions. Preserving the source structure, or adding explicit references between corresponding sections, can make the relationship clearer.

Annotations also make layout relevant. Ruby is a small annotation associated with base characters, often used to show pronunciation or provide explanatory material. In vertical writing, ruby is typically placed to the right of the base characters; in horizontal writing, it is typically placed immediately above them, according to the Japanese EPUB layout draft. Flattening the text without preserving the association can leave a reader unsure which characters the annotation explains.

A literary or language-learning text may rely on that association for reading. If an annotation is detached and placed at the end of a paragraph, the characters may still be present, but the reader has lost the relationship the layout communicates. A translator should identify whether the annotation gives pronunciation, an explanation, or another kind of reading support before deciding how to represent it in the target format.

Mixed-direction text can also make source layout important. A vertical page with horizontal captions, tables, or boxed passages contains several local reading orders. A reference copy that preserves those regions can help reviewers match each piece of translated text to the corresponding source element.

JLReq’s scope includes headings, illustrations, tables, notes, and reference marks, not only character recognition. That broader scope helps explain why a high-quality translation workflow includes both language review and page review. The issue isn’t simply whether every character was recognized; it’s whether a reader can still tell what each piece of text refers to.

For general reading, however, a faithful visual reproduction isn’t always the clearest choice. A report translated for colleagues who need its findings may be easier to use as a conventional horizontal document. The translated page can retain headings, table relationships, notes, and figure references while changing the direction of the main prose.

A useful decision is to ask what must remain stable: the source’s visual form, the connection between text and page elements, or the information itself. A project may need all three, but it may need only the latter two. Agreeing on that goal before translation prevents teams from spending time imitating a layout that the reader doesn’t need.

For practical guidance on translating multi-column layouts without reflow chaos, focus on preserving the relationship between regions as text changes length. Japanese tategaki adds direction-specific decisions, but the central question is the same: can the reader still follow the document?

When vertical layout isn’t needed¶

A translated document doesn’t always need to preserve its original writing direction. If readers need the content in English rather than a bilingual page-by-page reference, standard horizontal paragraphs may be clearer. The source layout still needs careful inspection, but the target can use a different arrangement.

A clean horizontal version can be the better choice for material that will be quoted, searched, read aloud, or reused in another document. The team can retain the original headings, notes, table labels, and figure references without forcing English into narrow vertical columns. A short note in the deliverable can also explain that the translated version uses a different direction from the Japanese source.

Reformatting doesn’t mean stripping out structure. Keep section boundaries, lists, table relationships, captions, and references visible. If a figure appears beside a vertical paragraph in the source, place the translated caption with that figure. If a note points to a particular passage, keep the note connected to that passage even when the page is rebuilt.

Choose a different output format when the original layout is too damaged or too complex to reproduce reliably. A scan with unclear text order, handwriting, or intricate regional relationships may need a human reviewer to establish the content before layout work continues. If the final reader only needs an accurate, legible translation, a structured text document can be more useful than an uncertain facsimile.

A table comparing formats can help a project team agree on the deliverable:

The table isn’t a ranking. The right format depends on whether readers need to compare the source, follow the original design, or read the translated material with minimal effort.

For teams that mostly handle translated forms and certificates, translating a certificate while keeping its original layout covers a related question: how much of a source document’s appearance should remain visible in the target.

A common mistake is to treat “preserve layout” as “copy every line break.” Japanese line composition includes constraints about characters that shouldn’t start a line, characters that shouldn’t end one, and sequences that should stay together. The Japanese EPUB layout draft describes these constraints. A line break that looks like a paragraph boundary in extracted text may therefore be a composition decision rather than a meaningful break in the content.

When translating into English, keep genuine paragraph and section boundaries, but don’t automatically reproduce every source line ending. English line lengths and paragraph flow work differently. Review the source to distinguish content structure from typesetting, then choose target-language line breaks that make the translated text easy to follow.

Common mistakes and how to catch them¶

Most problems with vertical Japanese translation start when someone assumes that a page has one direction, one text stream, or one correct target layout. A few deliberate checks can prevent those assumptions from turning into missing text, scrambled reading order, or detached annotations.

Treating vertical text as rotated horizontal text¶

Rotating an image can make the page easier for a person to inspect, but it doesn’t necessarily teach an OCR system how the Japanese columns are organized. Character grouping and reading order can still be wrong after rotation. The EasyOCR community discussion describes both column grouping problems and a rotate-and-detect workaround that didn’t fully solve recognition (issue discussion, comments dated 2022).

Check the extracted sequence against the page’s actual column order. If the system has combined characters from several columns into horizontal strings, correct the regions or use a workflow that can handle vertical writing. Don’t translate a questionable extraction and expect the translator to repair the layout problem by intuition alone.

Assuming every region follows the main direction¶

A vertical document can contain horizontal running heads, page numbers, captions, table entries, and boxed paragraphs. The historical EPUB draft identifies these as examples of local horizontal content in an otherwise vertical book (Japanese layout draft, 2010). Treating the entire page as one vertical stream can move a caption into the main text or put a number in the wrong sequence.

Mark local regions before extracting text. For each block, note whether it follows the page’s main direction and what it belongs to. A page number, for example, isn’t part of a sentence just because it sits near the last line of one.

Flattening ruby annotations¶

Text extraction can keep the visible annotation but lose the connection between the annotation and the base characters. In vertical Japanese, ruby is typically placed to the right of its base text; in horizontal Japanese, it is typically placed above it (Japanese EPUB layout draft). Once that positioning disappears, a reviewer may not know which pronunciation or explanation belongs to which word.

Preserve the annotation relationship in the editable source or represent it clearly in the translated output. A note beside the correct passage may work better than trying to mimic small ruby type in an English paragraph. The important decision is whether the annotation still serves the reader and remains attached to the right text.

Copying every source line break¶

Japanese line-breaking rules include characters that may not begin or end a line and sequences that should stay together, as described in the Japanese layout draft. A source line break can reflect those composition rules rather than a new sentence or paragraph. Copying every break into a translated English document can create choppy paragraphs or make relationships between phrases harder to follow.

Separate structural boundaries from visual line endings. Keep a genuine paragraph boundary or heading, but reflow ordinary text where English needs more room. Then check that any notes, references, or illustrations still point to the right content.

Trusting a model’s general OCR ability¶

A tool that reads horizontal Japanese well may still perform differently on vertical material. A 2026 LREC evaluation reports worse vertical-text reading performance than horizontal-text performance for the models and datasets the paper evaluated. The finding isn’t a universal scorecard, but it is a reason to test vertical pages as their own case.

Choose test pages that reflect the real document: include a page with several columns, a page with mixed text directions, and a page with annotations or tables if those occur in the file. Compare recognized text and its order with the source before processing the full document. A successful test on a simple page doesn’t establish that complex pages will be handled correctly.

Checking words but not page relationships¶

A text-only review can catch mistranslations while missing a caption under the wrong figure, a table entry in the wrong row, or a note detached from its reference mark. JLReq includes these broader composition topics alongside text composition (W3C JLReq). Check the translated page as a page, not only as a sequence of sentences.

A practical review sheet can pair each source block with its translated block, note its location, and record the reviewer’s check. The sheet doesn’t replace visual comparison; it makes the comparison easier to repeat and less dependent on remembering the order of a complex page.

Rebuilding without agreeing on the goal¶

A team can spend time reproducing a vertical layout only to discover that readers wanted a clear English reading copy. The reverse can happen too: reflowing a document may remove the structure needed for language study or source comparison. Decide at the start whether the deliverable is a visual reference, a bilingual comparison, or a readable English version.

A short project brief should state which page elements need to remain connected, whether the output should imitate the source direction, and how annotations will be handled. Those choices give translators, reviewers, and layout staff the same standard for deciding what to preserve.

FAQ¶

How do you translate a Japanese document written vertically?¶

First identify the page’s reading order and its separate text regions, then extract and translate the text while preserving the relationship between columns, annotations, and illustrations. Review the translated file against the original page rather than relying only on extracted text.

Why does OCR struggle with vertical Japanese text?¶

OCR systems may group nearby characters into horizontal lines, even when the page has vertical columns. Correct recognition also depends on column order and mixed-direction regions, so rotating the image alone may not fix the result.

How should Latin text and numbers be handled in Japanese vertical writing?¶

Check how each sequence is set in the original: individual upright characters, a sequence rotated clockwise, or a short horizontal sequence centered within a vertical line. Preserve the intended treatment instead of normalizing every Latin string the same way.

What should translators check when converting Japanese tategaki to English?¶

Check reading order, column sequence, mixed horizontal and vertical regions, line breaks, tables, headings, and the placement of ruby annotations. Compare names, numbers, and references with the original page.

How do ruby and furigana affect Japanese vertical text translation?¶

Ruby is annotation attached to base characters, often giving pronunciation or explanatory information. In vertical writing it is typically placed to the right of the base text, so extraction must preserve which characters each annotation belongs to.