Which Word Translation Add-in Is Best in 2026? Comparing Word, DeepL, Google, WPS and TopCopilot

Summary

This article breaks “which is best” into four testable questions: where does the translation end up; what level of formatting is retained; how can quality be evaluated without a standardized test; and how do data and costs affect selection? Choose the deliverable first, then the tool. Bilingual writeback necessarily changes pagination; cross-product quality rankings lack comparability; and the location of an interface does not establish whether data leaves the device.

1. Research design: turning “which is best” into testable questions

The question is difficult because it combines three goals: understanding, producing a deliverable, and continuing to edit. Each places different demands on a tool. Selection translation suffices for reading a passage; file-level translation is more direct for a separate translated copy; paragraph-by-paragraph comparison, consistent terminology and continued formatting in the original require access to the document object model.

We use four questions:

• RQ1 Processing path: Which steps from source file to usable deliverable are automated, and which remain manual?
• RQ2 Structural boundaries: Does “preserve formatting” mean fonts and paragraph relationships, or semantic objects such as tables of contents, fields and tracked changes?
• RQ3 Quality evaluation: How can a low-cost trial determine usability across different engines, language pairs and domains?
• RQ4 Constraints: How can platform, licensing, data flows and total cost rule out unsuitable options early?

Method: The primary evidence is public official documentation, help centers, pricing and privacy terms, supplemented by our team’s operational records. Sources were checked through November 12, 2026. This is not a controlled same-day test of all five products, nor a quality ranking of engines on a common corpus.

Evidence levels: Sources fall into four categories: A, official documentation (verifiable mechanisms and limits); B, vendor product pages and release notes (claims not independently verified); C, individual cases or operation records (limited samples); D, our inferences (judgments based on document models and engineering experience that require reader testing). Unlabeled descriptions of mechanisms are mostly A/B; explanations of why are mostly D.

What this article does not do: It does not compare BLEU/COMET scores for a particular language pair, claim that a product is “best,” generalize from one success or failure, or provide legal or compliance advice.

2. Technical paths: where does the translation end up?

The five products are often grouped as “Word translation add-ins,” but they operate on different units. By output location, there are four paths; the products do not map one-to-one to them: Word and DeepL offer both T1 and T2, WPS follows T4, and TopCopilot’s desktop add-in follows T3 while its web version follows T2.

Click to enlarge

The diagram shows only the two most common deliverables; the four technical paths differ in more detail. The key distinction between T1 and T2 is whether changes stay in the current document: T1 replaces selected text, normally leaving outside formatting and structure untouched; T2 sends the whole file to a service and returns a new file, leaving the original unchanged. T3 locates and writes back to paragraphs, table cells and graphical objects in the current document. It supports translation-only writeback or bilingual writeback, which retains the original and inserts translations paragraph by paragraph. T4 works similarly to T1/T2, but its capabilities and entitlements are tied to the WPS suite.

This is why describing TopCopilot only as “bilingual writeback” understates it: Its desktop version offers both translation-only mode for final delivery and bilingual mode for review, while its web version also supports file-level translation. It does not fit solely in the T3 box. See section 3.5.

An easily overlooked implication: When the output is a new file, interleaving original and translation paragraph by paragraph requires extra work rather than being the default. When the current document is modified, the user must back up the original first to preserve it. The deliverable determines the work after translation, often more than engine differences do.

Engine quality still matters. But for document delivery it is usually not the only bottleneck: even a fluent translation of a 50-page manual can require hours of cleanup if tables shift, contents page numbers break or terminology is inconsistent.

3. Analysis of the five tools

Each analysis covers mechanism, what can be automated, what remains unsolved, common misunderstandings and suitable conditions. Versions, plans and interfaces can change; check official pages and your account before purchase.

3.1 Word built-in translation: no installation, but large documents can fail

Mechanism: Word’s Review → Translate menu offers selection and document options. Preview a selection before inserting or replacing it; full-document translation produces a new translated document and retains the original. Microsoft also specifies eligible licensing, internet access and connected experiences as requirements; Microsoft 365 operated by 21Vianet does not offer this feature. See Microsoft support.

What it can automate: Preview and replacement of a passage; generation of a separate full translation; no additional installation or translation-service registration.

What it does not solve: Stability for large documents and long selections: users are more likely to encounter unavailable services, timeouts, oversized content or insertion failures. It also does not address interleaved paragraph-by-paragraph bilingual layout, glossaries, consistent translated font/size/color or resuming interrupted work. Confusing selection replacement with whole-document bilingual writeback is another common mistake.

Why large documents are prone to errors: Word’s built-in translation relies on an online service; full-document translation is not local. Longer documents and more tables, images and fields increase request size and the risk of timeouts, service errors or writeback failures. In our 129-page thesis case, both document and selection translation failed (C, n=1). This does not mean every large document fails, but large documents are a higher-risk scenario. If using the built-in feature, try 5–10 pages first and confirm service availability before processing the whole file.

Common misunderstandings: Built into Word does not mean free or offline. It saves installation, not post-translation cleanup. Another mistake is assuming every error is a network problem: long or large files are common triggers, while licensing, connected experiences and document complexity can also contribute.

Suitable conditions: You already use Word, occasionally read foreign text or translate a few paragraphs, work with short or medium documents, and accept a new copy. For documents hundreds of pages long, built-in full-document translation is not recommended as the primary approach.

Evidence and example: See Microsoft’s official support page for mechanisms and limits (A). In our earlier 129-page thesis comparison, both Word full-document and selection translation failed in that environment (C, n=1). This demonstrates that long documents can fail, but cannot establish failure rates for all documents. See the 129-page thesis comparison.

3.2 DeepL for Word: cross-platform use and terminology management, with payment limits to consider

Mechanism: DeepL offers a Word add-in. The sidebar translates selected text, with replacement after preview; full-document translation follows a separate path and creates a translated copy. Official help lists Windows, Mac and Word for the web, subject to Office version compatibility. Paid plans offer glossaries maintained on DeepL’s website and selectable in the add-in. See DeepL official help.

What it can automate: Cross-platform sidebar preview and replacement; full translation as a separate file; terminology consistency on paid plans; Mac and Word web users need not switch to a Windows native add-in.

What it does not solve: Inserting paragraph-by-paragraph bilingual text in the original; unifying translated font, size and color; final verification of complex tables, fields and tracked changes; configuring glossaries entirely within Office.

Common misunderstandings: A free account’s text allowance does not mean free full-document translation; that requires an appropriate DeepL Translator Pro subscription. Sidebar character limits, file allowances and size limits use different units and cannot be represented by one monthly figure. A brand’s reputation also does not guarantee superiority for your language pair and domain; without comparable testing, its name alone is insufficient. See also Pricing and Privacy Policy.

Suitable conditions: Mac or Word web users who accept a separate translation, need a glossary, are willing to pay and accept differences between free and paid terms.

Evidence: DeepL official help, pricing and privacy policy (A/B). We did not test DeepL engine quality under controlled comparable conditions.

3.3 Google Translate: convenient browser upload, outside the Word editing workflow

Mechanism: Google has two easily confused entry points: uploading documents on Google Translate and Translate document in Google Docs. The former lets you download the result; the latter creates a translated copy in Google Docs. Neither is an official add-in installed in Microsoft Word. See Google Translate help and Google Docs help.

What it can automate: Upload and download in a browser; support for formats including .docx, .pdf, .pptx and .xlsx; a free option for trying it; Google Docs users can continue sharing and editing the copy in their collaboration environment.

What it does not solve: Direct editing or paragraph comparison inside Word; translation of text in images and scanned PDFs; file size and page limits; network availability; format conversion when editing Office files in Google Docs.

Common misunderstandings: Image translation on Google’s website or Lens is not the same as translating images embedded in Word. The free website and Cloud Translation API have separate pricing and data terms; do not treat one as evidence for the other. A manual with many illustrations and little text can still hit the file size limit first.

Suitable conditions: Occasional files, cross-device access, acceptance of a downloaded translation and a reliable connection to Google services.

Evidence: Official Google Translate and Google Docs help (A). Consult current help pages for file size and page limits.

3.4 WPS built-in translation: convenient for WPS users, but not reusable across suites

Mechanism: WPS’s international site describes document translation, layout preservation and bilingual reading. It is a WPS suite feature, not a Microsoft Word add-in. Features, entry points and quotas differ across the Chinese and international editions and membership plans; the international feature list cannot simply be applied to a Chinese account. See WPS official introduction.

What it can automate: Document translation within WPS; low learning and switching costs for existing users; existing membership benefits may cover modest translation needs.

What it does not solve: Installation in Microsoft Word; sharing translation settings and workflows between Office and WPS; consistent entitlements across versions. Teams editing the final result in Word must still handle switching software and transferring files.

Common misunderstandings: WPS membership does not imply unlimited translation; free pages mentioned in old articles are not necessarily current entitlements. A local client interface does not mean inference happens locally. Check the feature’s data processing terms.

Suitable conditions: Primarily using WPS, with sufficient existing membership benefits and no need to replicate the workflow in Microsoft Word.

Evidence: WPS official introduction (B); check the client and account page for actual entitlements (A).

3.5 TopCopilot: translation-only and bilingual writeback across desktop and web

Mechanism: TopCopilot has two product paths. Its desktop add-in is a native Windows Office/WPS add-in offering whole-document, selection and selected-text translation in Word and WPS Writer. Full-document translation processes paragraphs and offers two writeback modes: Translation only (writes only the translation, suitable for final delivery) and Bilingual (retains the original and inserts translations paragraph by paragraph for close reading and review). Font, size, color and bold can be set; glossaries and custom prompts specify terminology and expression. The web version uploads .docx and downloads a separate translated file, suitable for Mac, occasional work or users who prefer not to install an add-in. Both share an account and credits. See Product overview, online translation and Change log.

What it can automate:

• Two writeback modes: Translation-only for a monolingual final copy; bilingual for paragraph-by-paragraph comparison. Bilingual output is not the only option.
• Three entry points: Whole document, selection and selected text. The latter two suit a chapter or paragraph, or reading before deciding whether to write back.
• Style control: Set translated font, size, color and bold. By default these follow the original; color can distinguish languages.
• Terminology and requirements: A glossary standardizes names and model numbers; a custom prompt specifies tone, audience and formatting and can be reused across project documents.
• Long documents: Paragraph-based processing without first splitting files; parallel windows, resumption from translated paragraphs after a pause and reuse of repeated short sentences within a task.
• Multiple hosts: One installer covers Word, Excel, PowerPoint and WPS Writer, Spreadsheets and Presentation. Translation of images inside documents is supported in Word/WPS Writer and Excel/WPS Spreadsheets, but not yet PowerPoint/WPS Presentation.
• Web version: Upload .docx to generate a separate translation on Mac or mobile, without installing the Windows add-in.

Why its coverage is broader: Among the five tools in our path table, TopCopilot alone covers T1 (selection/selected text), T2 (web file translation) and T3 (desktop writeback). Competitors usually cover part of this: Word is convenient to access but large documents can fail and it lacks glossaries, style controls and bilingual writeback; DeepL offers cross-platform use and glossaries but returns a separate full translation, leaving bilingual arrangement and translated fonts to other work; Google excels at uploading and downloading but is outside Word; WPS’s built-in tool is limited to WPS. TopCopilot combines translation-only and bilingual modes, styles, terminology, long documents, images and multiple hosts within one account and workflow.

What it does not solve: It does not promise completely unchanged formatting. Complex tables, equations, tables of contents, cross-references, tracked changes, text boxes and mixed text and graphics still need manual review. Bilingual output increases length and changes pagination. The desktop add-in supports Windows only. The web version returns a separate file, does not write back to the original and is not equivalent to translating images within documents. The personal edition requires internet access and sends text to the cloud; on-premises deployment does not necessarily have the same capabilities as the cloud version.

Common misunderstandings: Confusing translation-only and bilingual modes; confusing the web version with the desktop add-in; interpreting “the service does not persist document text” as “data never leaves the device”; assuming an enterprise on-premises deployment is an offline feature automatically included in a personal subscription. Image translation also sends image data off-device, which is easily overlooked.

Suitable conditions: For final translation, bilingual review, terminology consistency or image translation in Word/WPS on Windows, use the desktop add-in. For Mac, occasional use or a separate translated file without installation, use the web version. If text cannot leave the intranet, evaluate enterprise on-premises deployment. Personal plans require internet; confirm enterprise capabilities separately.

Evidence: Product and online translation pages, release notes and privacy policy (A/B); image translation and bilingual writeback screenshots are operation records (C). This article does not rank its engine quality against other products.

4. What does “preserve formatting” actually preserve?

“Preserve formatting” is among the most misused selection claims. At minimum, separate it into three levels:

L1 and L2 ask whether it looks like the original; L3 asks whether the document still works as before. Most claims cover the first two, but the third often causes the most rework.

In OOXML, a passage is not a single string. It consists of multiple <w:r> (runs), each with its own font, language tags and formatting properties. Spell checking, language tags and partial bold can split a sentence into multiple runs. Fields store code separately from cached results: tables of contents, page numbers and cross-references are examples. Translating the displayed result does not update the field. Tracked changes record insertions and deletions, and comments are separate annotation objects; they are not ordinary body text and cannot be treated as simple selection and replacement.

Three conclusions follow:

1. “Preserve formatting” is a layered promise, not a Boolean. A responsible claim says which layers are preserved and which objects need review, rather than promising an identical result.
2. Bilingual writeback necessarily changes pagination. Inserting translations between original paragraphs increases length and changes contents page numbers, cross-references and pagination. Requiring both bilingual comparison and an exactly unchanged page count is logically inconsistent. Bilingual output suits reading and review; choose translation-only before starting a monolingual final copy.
3. Non-text objects need separate handling. Chart titles, axes, category labels, SmartArt, equations, text boxes, headers, footers, footnotes and pixel text inside images are not necessarily in the main document’s text flow. Treating them as ordinary paragraphs can cause omissions or displacement.

Click to enlarge

Image translation is a separate capability: it handles pixel text inside images, not the document’s text layer. The next two images record one operation: select an English chart in Word, open image translation in the sidebar and replace the image with the translated result.

Click to enlarge

Click to enlarge

For detailed steps, batch mode and switching between original and translated images, see Image translation tutorial.

5. Translation quality: reproducible evaluation instead of cross-product rankings

Cross-product translation quality rankings are difficult to trust for several reasons:

• Engines, versions and prompts differ; the same product also performs differently across language pairs and domains.
• Document translation is not single-sentence translation. Context, consistent terminology, paragraph boundaries, numbers and units all affect usability.
• Evaluation criteria differ: some prioritize fluency, others terminology or formatting. Scores are not comparable without a common test set and scoring rules.
• Vendor demos generally select favorable examples; one success or failure is not a general conclusion.

Instead of ranking translation quality, we recommend an evaluation protocol that can be completed within 90 minutes. It does not replace large-scale testing but can help determine whether a tool is suitable for your document.

Test set: Select 40 passages from your actual document, stratified into four groups of 10:

Scoring: Two reviewers independently score blindly without seeing the brand. Each passage receives 2 = ready to use, 1 = usable with minor edits, or 0 = semantic error or rewrite needed. If scores differ by two levels, a third reviewer resolves the discrepancy. Record the disagreement rate; a high rate means the criteria need clarification.Critical errors tracked separately: Any critical error in numbers or units, negation or conditions, entity names or safety steps gives the passage a zero regardless of fluency.

Pass criteria: We suggest at least 80% of passages score 2 and zero critical errors. Otherwise adjust the glossary or change tools before committing to extensive manual repair. This is a cost-conscious starting point, not an industry standard; adjust it to document risk.

Record: Retain each source passage, translation, score, error type and editing time. Classify at least omissions, additions, mistranslations, terminology, numbers/units, negation/logic, style and formatting. This produces a comparable table rather than just an impression.

6. Terminology and instructions: deterministic and probabilistic constraints

Glossaries and custom prompts are often confused, but work differently.

• Glossaries A glossary is closer to a deterministic constraint: given a source term and target translation, it is expected to use the agreed output on a match. It suits product names, model numbers, brands, departments and fixed technical terms. Inflection, capitalization, multiword segmentation, partial matches and context-dependent meanings can all make a match incorrect. Glossaries improve consistency but cannot guarantee grammar or context.
• Custom prompts A custom prompt is closer to a probabilistic constraint: natural-language instructions specify tone, audience, purpose and format. It can improve style and expression but is not an auditable rule; the model may follow it in some paragraphs and ignore it in others. It suits soft requirements such as “translate in a manual’s tone” or “retain step numbers,” but should not be the sole safeguard for safety-critical wording.

A practical check: prepare 20 terms that must be consistent (five multiword terms, five model numbers or abbreviations, five easily confused terms and five items not to translate). After translation, count exact matches, incorrect variants and omissions. Then spot-check each of the ten prompt requirements. If glossary match rates are low, improve the glossary first; if prompt compliance is low, simplify the instructions or change tools. See Glossary tutorial and Bilingual translation tutorial.

7. Long documents: failure modes and recoverability

A long document is not simply a short document multiplied by its page count. Risks compound:

• Segmentation boundaries: Paragraph-by-paragraph translation loses cross-paragraph context; pronouns, ellipses and terminology can drift between passages.
• Context window: Sending more at once can make a model lose constraints later in the text; splitting more finely makes cross-paragraph consistency harder.
• Repeated content: Repeated headers, disclaimers and table headings consume allowances; products differ in whether they reuse translations.
• Partial failure: Network interruptions, depleted credits, rate limits and host stalls can stop a task midway. Whether completed translations were written back and whether work can resume matters more than total speed.
• Host state: Editing during scanning or writeback can shift paragraph anchors; bilingual insertion also changes later paragraph positions.

Our previous 129-page thesis case (C, n=1) showed Word’s full-document and selection translation failing while TopCopilot completed the task in that environment. It supports the conclusion that long documents need dedicated design, but not that all hundred-page documents fail similarly or any general success rate.

For long documents, evaluate recoverability rather than merely whether translation can finish:

8. Data boundaries: four questions rather than a local/cloud label

A button inside Word does not show that data stays on the device, and a server not saving data does not mean it never leaves the device. For sensitive documents, ask at least four separate questions:

1. What is sent: A selected passage, the full text, the entire file, or images/scans?
2. Who processes it: The direct provider or third-party models/subprocessors? Do free and paid versions follow the same path?
3. Retention and training: How long is it stored, where, is it used for training, and how is it deleted?
4. Legal and organizational constraints: Does it cross borders, meet the organization’s intranet, security and confidentiality requirements, and fall within contracts and data processing agreements?

Common pitfalls:

• “Document text is not stored” is a retention statement, not a claim of local processing. Text may still pass through servers and third-party models, which have their own terms.
• Images, glossaries and prompts are data too. Image translation sends images out; glossaries may contain undisclosed product or customer names, and prompts may include business context. These are usually account data, distinct from translated document text.
• Free and paid versions may have different terms. Do not apply one version’s privacy commitments to another.
• Clarify the scope of on-premises deployment. Ask who manages the model, servers, access and operations, and whether local models match cloud quality, language coverage and image capabilities.

This article does not provide a compliance conclusion. For contracts, financial, medical or tender documents, the organization’s information security and legal teams should assess their own rules; qualified professionals should review any tool’s output. For TopCopilot account data and retention, see Privacy Policy; for enterprise on-premises deployment conditions, see Enterprise overview.

9. Cost: replace subscription price with cost per deliverable document

Comparing monthly fees alone can be misleading. A more realistic measure is total cost per deliverable document:

C = subscription/quota cost + overage + human review hours × hourly rate + layout rework hours × hourly rate + switching/IT costs

Human time is usually the largest component. Consider an explicitly hypothetical example: a 20,000-word document reviewed at 1,000 words per hour and ¥100 per hour costs 20 hours and ¥2,000 for review alone. At 500 words per hour, the cost doubles. At this scale, whether the subscription is ¥10 or ¥30 is not decisive. Conversely, a tool that saves time finding, copying and comparing passages may save far more than the monthly fee difference.

Quota units cannot be compared directly: characters, credits, pages and files represent different amounts of work. The most reliable approach is a small trial with a typical document, recording consumption per 1,000 characters or 10 pages, then calculating cost per document. TopCopilot’s published list prices (Pro ¥10/month with 10,000 credits; Pro Max ¥30/month with 30,000 credits; check the account page) are a starting point, but compare the credits and human hours required to complete your document. Image translation is charged per image (the interface shows about 450 credits per image; check your account) and should be estimated separately from text. See all plans Pricing page.

10. Decision rules: eliminate options before choosing

Instead of asking which is best, eliminate options based on your constraints. This table compares scenarios:

Three decision principles:

1. Choose the deliverable first, then the tool. T2 is more direct for a separate translation; T3 matters when editing continues in the original; T1 is enough to read a passage.
2. Inventory structure before discussing format preservation. Does the document contain a table of contents, cross-references, tracked changes, text boxes, charts or scanned images? These determine the review checklist and the credibility of lossless-format promises.
3. Define data boundaries before comparing price. Online personal services are unsuitable when material must stay on the intranet, however cheap. Compare subscriptions and quotas only when cloud processing is acceptable.

11. Research limitations and conclusions

Limitations:

• This article relies primarily on public documentation. It did not conduct controlled same-day tests of the five products on the same document and language pair; it cannot establish engine quality rankings.
• Product features, plans, interfaces and terms change. Sources were checked on November 12, 2026; verify them again before purchasing.
• Vendor documentation is self-reported and some claims have not been independently verified. Evidence grading identifies this but cannot remove all uncertainty.
• The 129-page thesis and image translation examples are single observations (n=1), not general success rates. The 40-passage protocol is stratified sampling from one document and cannot replace statistical evaluation across documents and fields.
• This article is not legal, compliance or procurement advice.

If you remember only three things:

1. Where the translation ends up is the first determinant of delivery efficiency, ahead of brand. A separate translated file, writeback in the original and paragraph-by-paragraph bilingual output are different workflows.
2. Assess formatting preservation at three levels. Visual layout, structure and semantic objects: bilingual writeback necessarily changes pagination; update contents and cross-references.
3. Use your own quality evaluation, not a ranking. A stratified 40-passage test, two blind reviewers and separate tracking of critical errors are more reliable than a feeling that one translation reads better.

Conclusion: Word built-in suits existing Word users who occasionally need comprehension and accept a new copy; DeepL suits cross-platform users who want a separate file and will pay for glossaries; Google suits occasional upload/download across devices; WPS suits those primarily using WPS with sufficient existing benefits. TopCopilot covers desktop writeback (translation-only or bilingual) and a separate web translation, fitting documents that need further editing, comparison and delivery; its desktop add-in is Windows-only. This is a match between tools and deliverables, not a ranking. Define the deliverable, structural boundaries and data constraints before choosing a tool.

Appendix: 90-minute evaluation checklist

Take a representative document with headings, tables, numbering, technical terms and at least one image containing text. Check the following ten items and record pass/partial/fail plus an example for each.

Suggested record fields: document name, tool/version, language pair, pages/word count, start and finish times, quota consumed, failure points, hours of manual editing, critical errors and conclusion (usable/needs edits/unusable). This record reflects your actual costs better than any overall score.