What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
To translate a whole book with an LLM without losing context, split the text at its own structure (chapters, sections, paragraphs) rather than at a fixed character count. Give every segment what a human translator keeps open while working: a brief, a glossary of names and terms, the translated text that came before, and an ID that ties each translated paragraph back to its source. The sources reviewed do not establish a universal chunk size or overlap. The practical target is the largest structure-aware segment your chosen model handles reliably in your language pair and genre, and you confirm that by checking the output.
Why isolated chunks lose the book
Cutting a manuscript into equal-sized slices is the simplest approach, and it fails in predictable ways. Slices cut through dialogue exchanges, scene breaks and footnotes, and each slice is translated as though it were the whole text. The examples below illustrate the failure pattern; they are not measured counts from a test.
As an Amazon Associate I earn from qualifying purchases.
- Names and titles. A character introduced as “Captain Reyes” in chapter 2 may be rendered one way there and differently in chapter 14, because nothing told the model the earlier choice.
- Pronouns and references. A segment that opens with “She did not answer” loses its referent when the antecedent sits in the previous chunk.
- Invented or specialized terms. A term the author defines in one chapter is translated freshly each time it reappears.
- Cross-references. “As noted earlier” or “the plan from chapter 3” has nothing to point to if chapter 3 is not in view.
- Voice. A narrator’s clipped sentences come out smooth in one segment and stiff in the next, so the book reads as if several translators wrote it.
Why one request for the whole book is not the answer
A whole-book prompt avoids cutting, but two limits apply. First, input and output limits are set per model, and a translation can run longer than its source in the target language, so the output budget can bind even when the input fits. Check the provider’s documentation for the model you use rather than assuming a window size.
Free tools Windows power users keep installed
One-click scans. No signup required.
Second, an advertised context window is not a guarantee of quality. Wang et al., EMNLP 2025, introduced SEGALE, a method for evaluating long-document translation, and reported that many of the open-weight LLMs they evaluated did not translate book-length texts effectively at their reported maximum context lengths. Test any long window on your own book before relying on it.
#1 Best Overall
Three workflows compared
There are three practical strategies. They differ mainly in what each model request knows about the rest of the book.
| Question | (a) Whole document, single turn | (b) Independent chunks | (c) Source-primed, iterative multi-turn chunks |
|---|---|---|---|
| Does each request see source context? | Yes, the whole document, if it fits the budget | Only its own chunk | Yes: the source (in full or summarized) is provided first |
| Does it carry earlier translated text? | Not applicable; the book is produced in one pass | No | Yes: previously translated segments are kept in the conversation or prompt |
| Does it track terms and names? | Only if the glossary is included in the prompt | Only if a glossary is injected into each chunk | Yes: the glossary is updated and supplied as the book proceeds |
| Can source and target be aligned segment by segment? | Hard to map back to paragraphs unless output is returned with IDs | Yes, if each chunk keeps its IDs | Yes, with stable segment IDs and a mapping table |
| Can a run resume after failure? | A failure usually means rerunning the whole book | Yes, chunk by chunk | Yes, from the last saved segment |
| Main trade-off | Bounded by the model’s usable input and output size | Lowest setup effort, but context is lost | More pipeline work, and segments must be translated in sequence |
The workflow, step by step
- Start from structured text. Use an EPUB, or a clean text or Markdown export that keeps headings, paragraphs, lists, footnotes and section breaks. PDFs need an extraction step that preserves the same boundaries. If the book is a scan, OCR comes first; it is conditional, and OCR errors pass straight into the translation, so proof extracted names and numbers against the page images.
- Write a brief for the whole book. Include the genre, intended readers, target-language register, translation principles (for example, keeping the narrator’s sentence length or dialect markers), recurring names with chosen target forms, key terms with approved renderings, forms of address and how they shift with social rank, and anything that should stay untranslated.
- Segment at structural boundaries. Split by chapter, then section, then paragraph. Never split inside a sentence, and keep dialogue exchanges together where you can. The sources give no word or token count, so size each segment against the model’s limits and the output you actually get in your target language.
- Assign stable IDs. Use a format such as
ch03-s02-p014, and keep a mapping table with source ID, target ID, status and revision number. Never reuse an ID. - Build the context for each request. Send the brief; the glossary entries whose source terms appear in this segment; the source, in full if it fits your budget or as a summary if it does not; the last few translated segments; and the segment to translate. How many prior segments to include is a setting to test, not a fixed rule.
- Translate in order and save each output. Store every output with its segment ID, model name, date and prompt version. If a run stops, resume from the last saved ID instead of restarting.
- Update the glossary as terms appear. When a name or term is decided, record it once with its approved rendering, and make later segments use it.
Choosing a segment size
Use a decision process rather than a preset number:
- Treat the model’s documented limits as a ceiling, not a target. Stay well under the input and output limits so no translation is cut off.
- Let structure set the upper bound. If a chapter fits comfortably, keep it whole; if not, split at section breaks, then paragraphs.
- Adjust for text density. Dialogue-heavy fiction works with short turns grouped by scene. Technical or legal books carry many terms, so the glossary matters more than segment length.
- Decide how much source context you can afford. Full source fits only shorter works; for longer ones, a source summary plus the glossary is one option to test.
- Handle overlap with care. The sources do not establish whether overlapping segments help. If you add overlap, check the assembled text for duplicated passages.
- Test two sizes on representative chapters. Translate two or three chapters at two segment sizes, then compare name consistency, omissions, and a reviewer’s notes on voice. Choose the size that does better on those checks.
What the published studies show
The studies below are the strongest available evidence for context-aware translation. Each tests a specific setup, and none establishes a universal chunk size.
Source-primed multi-turn translation (Hu, Vamvas and Sennrich, 2025)
Hu, Vamvas, and Sennrich, in the Association for Computational Linguistics’ Findings of EMNLP 2025, tested a source-primed multi-turn approach: the whole source document is provided first, then segments are translated iteratively. The authors write:
“We empirically show this multi-turn method outperforms both translating entire documents in a single turn and translating each segment independently according to multiple automatic metrics in representative LLMs, establishing a strong baseline for document-level translation using LLMs.”
That result holds for the authors’ tested setup, automatic metrics and models. It does not show that every multi-turn system performs best.
Paragraph context in literary translation (Karpinska and Iyyer, WMT 2023)
Karpinska and Iyyer compared paragraph-level and sentence-level translation of literary text using GPT-3.5 (text-davinci-003), a model from 2023. In their human evaluation, paragraph-level methods produced fewer mistranslations, grammar errors and stylistic inconsistencies. The evaluation was labor-intensive: fluent speakers of the source and target languages supplied span-level error annotations and preference judgments, a total of 350 hours of annotation and analysis. The paper is also direct about the limits:
Recommended Free Tools
“With that said, critical errors still abound, including occasional content omissions, and a human translator’s intervention remains necessary to ensure that the author’s voice remains intact.”
Because the model predates the current generation, read this as evidence of the direction of the effect, not as a comparison of today’s models.
Long-document evaluation (Wang et al., EMNLP 2025)
Wang et al. introduced SEGALE to evaluate long-document translation. Their finding that many evaluated open-weight LLMs did not translate book-length texts effectively at their reported maximum context lengths is the reason to test long inputs directly. SEGALE is an evaluation method, not a quality certification.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Quality control and where human review fits
Automated checks catch structural problems but cannot judge voice. Run these after each chapter and again on the assembled book:
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →- Segment count and order. Every source ID should map to exactly one target ID, in sequence. A missing ID points to a dropped paragraph, the omission risk Karpinska and Iyyer describe.
- Cut-off outputs. Look for sentences that end mid-clause, and check the model’s stop reason. A translation far shorter than its source can signal truncation.
- Structure and formatting. Headings, lists, footnotes and emphasis should survive. Stray prompt text or explanations in the output mean the prompt leaked, and the segment must be rerun.
- Glossary compliance. Search the assembled text for each approved term and name, and flag any that appear in a second form.
- Late-chapter drift. If names or terms start shifting near the end, update the glossary and rerun the affected segments so the saved record matches the final text.
Human review is the final stage, not an optional extra. Have a qualified bilingual reader go through the assembled text, prioritizing ambiguous passages, dialogue, figurative language, and the opening and closing of each chapter. For literary work, the reviewer should check voice across the whole book, not only sentence accuracy. For publication, judge the complete translation as a document rather than from sentence-level scores alone. A review by a qualified bilingual or literary translator catches errors that automated checks miss, but it lowers risk; it does not make the translation error-free.
What to look for in workflow tools
Dedicated long-document tooling can handle the bookkeeping described above. The ContextWeaver documentation describes stable segments, evidence-backed terminology and entity context, revision records, validation, and Markdown and EPUB export. That describes the project’s capabilities; it is not an independent test of translation quality. When evaluating any tool, check for these functions:
- Stable segment IDs and a source-to-target mapping
- A glossary or entity store shared across all requests, which can be updated mid-book
- Saved revisions, so changes can be traced and reverted
- Resumption after a failed run
- Validation of omissions, order and formatting
- Export to the formats you need, with a point where a human reviewer can edit
For a single book, a spreadsheet for the mapping table and a short script for segmentation and saving can cover the same functions.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
