A document is not a wall of text.

Paste a paragraph into a translation box and you get a translated paragraph back. Simple. Now try the same with a 40-slide board pack: the text lives in titles, tables, speaker notes, SmartArt diagrams, and a chart legend, the slides are laid out right to left, and two images contain text baked into the pixels. Suddenly "translate this" is not a language problem. It is a document engineering problem wearing a language problem as a hat.

This lesson is about why documents are harder than text, because understanding the failure modes is the difference between a translated copy you can send to the board and an evening of manual cleanup.

Structure: text does not live where you think it lives

A Word file is not a wall of text. Internally it is a tree of paragraphs, and each paragraph is split into runs: fragments that share the same formatting. Bold one word in a sentence and that sentence becomes three runs. A naive tool that translates run by run feeds the model sentence fragments, and as lesson one explained, fragments without context translate badly. A good pipeline reassembles runs into sentences, translates the sentence, then maps the formatting back onto the translated words, deciding which word in the Arabic output deserves the bold that one English word carried.

Then there is everything that is not body text: tables (where each cell is its own little document), text boxes floating over the page, headers and footers, footnotes, slide speaker notes, and SmartArt, where text lives inside drawing objects several layers deep. Each surface has to be found, translated in context, and written back to exactly the same place. Miss one and the translated document ships with a stray English footer nobody noticed until the client did.

Length: translated text is a different size

Languages do not occupy the same space. Going from Arabic to English, text typically expands by roughly 20 to 40 percent; the compact Arabic root system packs meaning into fewer, shorter words than English needs. German famously expands from English; Chinese contracts dramatically. The layout you carefully designed for one language is now holding a different volume of text.

The consequences are mundane and infuriating: headlines wrap onto a second line and push everything down, table cells overflow, text boxes clip their last line invisibly, a slide that held five bullets now holds four and a half. A translation pipeline that ignores this hands you a document where every page needs manual nudging. One that handles it adjusts sizing and flow so the translated copy remains a finished document, not a repair project.

Images: text with no text in it

Organizations love putting text inside images: scanned stamps, infographics, a chart exported as PNG, a screenshot of a table pasted into a report. To a translation pipeline, an image contains no text at all, just pixels. Handling it takes three steps: OCR to read the words out of the pixels, translation, and then rendering the translated text back into the image, masking the original so the result looks like the picture was made in the target language in the first place. Skipping the third step gives you the common half-measure: a translated document with stubbornly untranslated pictures in it.

PDFs: printouts pretending to be documents

A PDF deserves its own paragraph of respect and suspicion. PDF is a print format. It does not store paragraphs, or even reliably store words; it stores instructions like "place these glyphs at these coordinates." Reading order, paragraph boundaries, table structure: all of it has to be reconstructed by the tool, essentially reverse-engineering the document from its printout. Two-column layouts, pull quotes, and tables make this genuinely hard, and scanned PDFs add OCR on top. When a translated PDF comes back looking right, an enormous amount of reconstruction happened quietly on your behalf.

Direction: when the whole page flips

Arabic to English is not just a language change, it is a direction change. Right-to-left layouts mirror: what was right-aligned becomes left-aligned, sidebars swap sides, table columns reverse order, and in presentations the shapes themselves sit in mirrored positions. A pipeline that translates the words but leaves an RTL skeleton produces documents that feel subtly wrong to every native reader of the target language, even if they cannot say why. Lesson five digs into direction and Arabic typography properly; for now, know that direction is a document-level property, not a font setting.

What this means in practice

When you evaluate any document translation tool, ignore the demo paragraph. Feed it your ugliest real file: the PPTX with SmartArt and speaker notes, the PDF with two columns and a scanned annex, the DOCX where legal bolded every third defined term. Then check the four failure points in order: did every text surface get translated, did the layout survive the length change, did images with text get re-rendered, and did the direction flip cleanly?

This whole problem class is what TranslateX treats as the core product: one-click document translation that produces a translated copy next to the original with fonts, tables, spacing, and slide structure preserved, and layout preservation doing the reconstruction work described above, including re-rendered images and mirrored PPTX shape positions when direction flips. The next lesson moves from the container to the words: keeping terminology consistent across thousands of segments.