LisanMultilingual
document translation.
T3 powers document intelligence within TranslateX: language, structure and a workflow for professional review.
Product capabilities, system design and an Arabic–English document evaluation.
LisanTranslate the language.
Preserve the document.
Translation quality and document structure assessed together, across Arabic and English workflows.
translation-quality score
structural-fidelity score
directions
Translation that keeps the document useful
T3 combines language generation with document processing inside TranslateX. It translates content together with the structure that gives it meaning: headings, paragraphs, tables, typography and reading direction.
The platform extends that engine with multilingual text and document workspaces, approved terminology, translation memory, professional review and API integration. Native and scanned documents follow appropriate preparation and reconstruction paths.
Product scope
A multilingual portfolio, document and structured-text formats, reviewer tools and customer-controlled deployment options.
Measured scope
Arabic-to-English and English-to-Arabic document evaluation across DOCX, PPTX and PDF, with DOCX predominating.
Reading the paper
LisanMultilingual by design.
Direction-aware throughout.
TranslateX supports a broad language portfolio for text and document workflows.
Languages in the product catalogue
Text and layout together
Language selection informs translation, font handling, alignment and reading direction. Document reconstruction adapts the output where the format supports it.
Configure for the workflow
Select the language pair, model configuration and terminology for the task. Script, document type and domain all shape the review requirements.
LisanMany formats.
One document engine.
Translate the content while preserving the structure needed to work with it.
Documents
Paragraphs, headings, lists, tables, headers and footers.
Presentations
Text boxes, diagrams, shapes and the relationships between them.
Spreadsheets
Text within the grid, with numeric values and identifiers kept in context.
Reports and scanned files
Extract or recognise content, recover layout and reconstruct an editable output where supported.
Preserve more than the words
Meaning, names and domain expressions
Tables, lists, hyperlinks and charts
Fonts, images, alignment and RTL/LTR layout
Read document objects and styles directly.
OCR recognises text; OLR reconstructs the page layout.
Available inputs and editable exports depend on the configured workflow. DOCX, PPTX and PDF appear in the scored evaluation; XLSX has a separate visual case. Other listed formats and OCR/OLR capabilities are platform scope, not additional benchmark results.
LisanFrom file analysis
to translated output.
Six stages connect language generation to the structure of the document.
Analyze file structure
Detect the file type and identify its structure.
Extract text and styles
Extract text and style information from body paragraphs, tables and all supported text blocks.
Translate text
Translate the extracted text using the selected translation model.
Apply font styles
Adjust font names and sizes when necessary, based on the target language and the text blocks.
Apply text direction
Apply right-to-left or left-to-right direction and mirror shape positions when necessary, based on the target language.
Export the translated file
Export the final document in a supported output format.
Conceptual illustration. Font adjustments and shape-position mirroring are applied when required; they are not universal transformations for every block.
Language and layout
Arabic typography and the transition between right-to-left and left-to-right scripts are central to the workflow. Text is processed together with style information so the output can retain its document function.
Model and output scope
The workflow supports several model families. Model selection and supported export formats define the operating configuration; the benchmark assesses the saved outputs.
LisanTranslate. Review.
Reuse what is approved.
A professional translation workflow brings language assets and human decisions into one place.
Files, languages and layout options
Configured engine and terminology
Source and target together
A recorded reviewer decision
Confirmed language assets
Glossaries
Upload, create and edit preferred terminology. Keep names, abbreviations and domain expressions consistent with the organisation’s language policy.
Translation memory
Reuse approved segments and project language preferences across recurring documents. Human-reviewed translations become reusable language assets.
CAT review
Compare source and target segments in a computer-assisted translation editor, revise the text and confirm the reviewed version.
Projects and versions
Organise documents, languages and processing status. Track the translated document’s history and preserve the link to the source.
Prepare the source → apply approved terminology → review the meaning and document layout → approve the version → return the reviewed output to its authorised workflow.
This is a workflow illustration, not a customer deployment claim. Roles, release controls and domain language assets are configured for each enterprise. The document benchmark measures output quality and structure separately from reviewer productivity.
LisanYour documents.
Your enterprise context.
A reference deployment architecture for controlled translation and professional review.
T3 translation, OCR and layout reconstruction
Approved terminology, memory and model configuration
Source-linked editing, approval and output history
Connect to existing work
Use APIs and asynchronous document-job callbacks to connect translation to document repositories, enterprise applications and review queues.
Keep administration explicit
Define identity integration, permission scopes, model endpoints, retention and residency for the selected environment. Verify these controls during deployment acceptance.
LisanDocument-level results.
Both language directions.
All records receive equal weight. Language direction follows the recorded source and target languages.
Adequacy
How well the meaning is preserved.
Fluency
How natural the target language reads.
Completeness
How fully content is carried across.
Results by recorded direction
| Measure | Arabic → English | English → Arabic | All records |
|---|---|---|---|
| Quality / 100 | 93.34 | 93.34 | 93.34 |
| Adequacy | 91.50 | 91.88 | 91.74 |
| Fluency | 93.80 | 93.10 | 93.36 |
| Completeness | 94.73 | 95.07 | 94.94 |
| Structure / 100 | 96.70 | 98.67 | 97.93 |
Mean quality rounds to 93.34/100 in both directions. The groups contain different documents; their means do not establish a statistically significant language-direction advantage. Structural fidelity uses applicable document checks and remains separate from language quality.
Figure 2. All bars use the same 0–100 scale. Overall quality averages stored file-level scores; it is not recomputed from the rounded dimension means shown here. Null structural criteria are excluded at the file level.
LisanA protocol readers
can inspect.
Descriptive summaries are recomputed from saved evaluation records. Translation and model judging are not rerun for this paper.
Corpus and language direction
The corpus contains saved document-level evaluations across the stated formats. Results are grouped by recorded source and target language. All records are marked as successfully translated.
Source and translated content
The evaluator compares source content with its translated output. Text extraction and scoring coverage depend on the file structure and the content available for assessment.
Language quality
A model judge rates adequacy, fluency and completeness from 0 to 100. DOCX and PPTX use segment-level judging; PDF uses holistic judging. The benchmark averages the stored file-level quality scores.
Structural checks
Where applicable, checks cover direction, alignment, font family and size, image counts, headers and footers, margins, textbox coordinates and estimated text containment. Available checks vary by format; null criteria are excluded from each file’s style score.
Aggregation
Stored file-level scores are averaged with equal record weight and displayed to two decimal places. The headline means use no word, page or segment weighting. Rounding already applied by the evaluator is retained.
Q is a stored document quality score.
n is the number of file records.
The evaluator rounds each file style score.
The paper averages the stored Sᵢ values.
Successful-translation status describes the recorded corpus. The scores assess its output content and structure; they do not make every output error-free or establish an independent acceptance rate.
LisanLanguage, layout
and operating scope.
The results measure language quality and document structure. Timing and deployment controls require separate assessment.
Interpreting the recorded timing
Most records reuse cached outputs and have zero stored translation time. Positive timings are recorded only for uncached runs. Zero does not mean instantaneous translation. These snapshots do not support a comparable end-to-end throughput claim.
Remaining output errors
Some terms are mistranslated and some source-language text remains untranslated. Alignment preservation is weaker in Arabic-to-English output. Targeted review is still required.
Corpus and scoring limits
These are internal automated scores, not independent validation or accuracy percentages. Records have equal weight even when document lengths or applicable structural checks differ.
Comparison limits
The following samples are illustrative and separate from the aggregate benchmark. Vendor labels follow supplied filenames. Individual outputs do not establish general comparative superiority.
Model and deployment boundaries
Model-family support and deployment flexibility describe the system. The benchmark does not test security, hosting boundaries, data residency or operational controls; these need evidence for the selected configuration.
Evaluate the workflow on representative documents
Inspect language meaning and document function together: terminology, remaining source text, column order, alignment, typography, tables and text containment. The checks most relevant to a use case depend on its file structures and translation direction.
The comparisons show actual supplied outputs from TranslateX, Google, Microsoft and DeepL. XLSX is illustrated separately from the DOCX, PPTX and PDF scoring scope.
LisanTwo columns.
One document workflow.
English → Arabic · Page 1 of the supplied DOCX outputs.
Column order
Inspect how the two-column structure reads in the target language.
Typography
Compare headings, paragraph styling and Arabic text appearance.
Tables
Inspect table placement, cell alignment and retained numeric values.
Original pages and translated outputs are shown in full. Use a larger zoom to inspect text and layout; these illustrations sit outside the aggregate benchmark.
LisanSupplied DOCX outputs.
English → Arabic · Page 1 · Visual examples, not a benchmark ranking.
Illustrative supplied outputs; engine labels follow the filenames. These examples are separate from the aggregate benchmark and do not establish general comparative superiority. Layout, clipping and overlap are retained.
LisanA slide, translated
as a visual document.
English → Arabic · Business strategy presentation · Slide 1.
Inspect the relationship between text, diagram labels, shapes and alignment. Illustrative outputs are separate from the aggregate benchmark; visual samples alone do not establish semantic accuracy.
LisanSupplied slide outputs.
English → Arabic · Business strategy presentation · Slide 1.
Illustrative supplied outputs; engine labels follow the filenames. These examples are separate from the aggregate benchmark and do not establish general comparative superiority. Layout, clipping and overlap are retained. DeepL is represented by the supplied screenshot.
LisanSpreadsheet content
within its grid.
English → Arabic · Inventory records · Worksheet A1:R22.
Column structure
Inspect headings and the ordering of the inventory fields.
Cell content
Compare translated text, identifiers and numeric data.
Readability
Inspect cell dimensions, alignment and text containment.
LisanSupplied spreadsheet outputs.
English → Arabic · Worksheet A1:R22 · Outside the stated benchmark formats.
Illustrative supplied outputs; engine labels follow the filenames. These examples are separate from the aggregate benchmark and do not establish general comparative superiority. Layout, clipping and overlap are retained. DeepL is represented by the supplied screenshot.
LisanA report across languages.
One representative page.
English → Arabic · WFP strategic plan report.
Page structure
Inspect paragraphs, headings and document furniture.
Writing direction
Compare the source layout with the Arabic output.
Content flow
Text can reflow between pages across different outputs.
Source PDF; TranslateX DOCX output. Matching page numbers may contain reflowed text. Full output pages are retained, and these illustrations are separate from the scored evaluation.
LisanThe same report.
Comparison outputs.
English → Arabic · Matching page numbers; text may reflow between outputs.
Illustrative supplied outputs; engine labels follow the filenames. These examples are separate from the aggregate benchmark and do not establish general comparative superiority. Layout, clipping and overlap are retained. Microsoft used Azure Document Translation. Matching page numbers may contain reflowed text.
LisanFrom measured results
to document review
The headline scores summarize saved automated evaluations. Each file receives equal weight, so the composition of the corpus matters when interpreting the result.
DOCX
Dominant format in the quantitative evaluation.
PPTX & PDF
Less extensively represented in the same averages.
XLSX
Illustrative comparisons only; outside the scored corpus.
The format imbalance matters: coverage does not imply equivalent validation across formats. Evaluation scope and scoring methods ↗
Aggregates checked against saved records
Recomputing the equal-record means gives 93.34/100 for quality and 97.93/100 for structure. The Arabic-to-English and English-to-Arabic groups retain their own documents and cannot establish a causal language-direction advantage.
Timing and output evidence stay separate
Cached outputs have zero stored translation time; uncached runs have positive timings. No speed advantage follows from those zeros. The comparison screenshots retain the supplied outputs, including clipping, reflow and overlap.
Quality and structural fidelity are scores out of 100, not accuracy percentages or a quantitative ranking against competing translation products.
Open the evaluation records ↗Inspect the document comparisons ↗