The two most common problems when AI translates long documents are dropped passages and inconsistent terminology. Dropped passages mostly happen because the document is too long and a single input exceeds the processing range; inconsistent terms happen because the AI re-guesses the translation every time it meets the same word. Both problems must be handled before translating, not fixed sentence by sentence afterwards.
Split first: cut by structure, not by character count
Split the document by chapter, subheading and paragraph, and translate only one complete section at a time. Choose split points at chapter boundaries, never in the middle of a sentence, a table or a list; once the cut context is lost, omissions and mistranslations follow. Before each section, give the AI one sentence of background: what kind of document this is, who it is for, and whether the translation should be formal or colloquial. After translating the sections, reassemble them in order and read through the joins between sections, focusing on whether sentences are missing at the start and end of sections and whether punctuation and numbering are disordered.
Build a glossary: fix the translations before starting
Before translating, pull out the key terms in the document: product names, people's names, place names, specialist words and abbreviations. Fix one translation for each term in a reference table, and send it to the AI together with the requirement of which words to leave untranslated in the original language. Product and brand names mostly stay in the original; people's names follow the established translation; place names follow the official translation. Require the AI to follow the table strictly, and when it meets a new term not in the table, mark it out rather than inventing a translation. When a long document is translated over several days, bring the same glossary every time, or the front and back will certainly drift apart.
How to check for dropped passages yourself
After reassembling the full text, do three quick checks: whether the paragraph counts match between source and translation; whether numbering and list items are complete and table rows are not missing; and whether all subheadings are present. If a passage is missing, do not just ask the AI to fill in the gap and finish; send the missing passage together with the two passages around it so it can fill it in with context. Text inside charts, footnotes, headers and footers is easiest to miss; before translating, confirm whether that content is in the input, and if not, extract it separately and translate it on its own.
For formal documents, the AI translation is only a draft
For formal documents such as contracts, papers and product manuals, the AI translation must be checked by a person afterwards on the key sentences: numbers, dates, amounts, liability clauses and operating steps, none of which can be trusted to the AI alone. Check the glossary, the source text and the translation against each other, and change only the translation, not the glossary. Deliver only after checking, and read the full translation once more before delivery to confirm it reads like one complete document rather than a pile of stitched-together paragraphs.
