Automating Translation Workflows with LLMs for Post-Editing and Transcreation

transperfect · Tech · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise localization and translation workflow automation

TransPerfect and the AWS localization team needed to reduce cost and turnaround time for translation, post-editing, and creative transcreation workflows while preserving quality, data security, auditability, and translation accuracy.

Constraints:

• AWS localization handles billions of words across many digital assets and languages
• translation workflows include asset handoff, preprocessing, machine translation, post-editing, quality review, and handback
• creative transcreation needs cultural nuance and historically resisted automation
• client data cannot be shared with model providers or used to improve base models
• translation output requires contextual grounding and hallucination protection

Architecture

TransPerfect integrated Amazon Bedrock into the GlobalLink translation management system. Existing localization workflows reuse translation memory first, route unrecycled segments through Amazon Translate machine translation, then use Bedrock-hosted LLMs for automated post-editing with style guides, approved examples, and error examples. Creative transcreation uses Claude or Amazon Nova Pro to propose multiple adapted translations for a human linguist to select, while Bedrock Guardrails provides truthfulness protections.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  globallink-translation-management-system[GlobalLink translation management system] --> translation-memory[Translation memory]
  globallink-translation-management-system --> amazon-translate[Amazon Translate]
  amazon-translate --> amazon-bedrock-llm-post-editor[Amazon Bedrock LLM post-editor]
  style-guides-and-approved-examples[Style guides and approved examples] --> amazon-bedrock-llm-post-editor
  amazon-bedrock-llm-post-editor --> amazon-bedrock-guardrails[Amazon Bedrock Guardrails]
  amazon-bedrock-llm-post-editor --> human-linguist-editor[Human linguist editor]
  bedrock-transcreation-model[Bedrock transcreation model] --> human-linguist-editor
  human-linguist-editor --> globallink-translation-management-system

Components

Data flow

A localization asset enters GlobalLink, where translation memory is applied first. Segments that cannot be recycled are machine-translated with Amazon Translate. The machine translation can then be post-edited by a Bedrock LLM using style guides, approved translation examples, and error examples. Depending on the workflow, the result is either sent to a linguist for lighter post-editing or applied in no-human-touch workflows; creative assets can instead be transcreated by Claude or Amazon Nova Pro with human selection in the editor.

Why this architecture

The architecture keeps existing localization workflow controls while adding LLM automation where it can reduce human post-editing and transcreation effort. Translation memory remains first because approved prior translations are safest to reuse; Bedrock is used for secure LLM post-editing and creative candidate generation, and human linguists stay in the loop where quality, nuance, or selection still matters.

Models used: Amazon Bedrock foundation models · Amazon Translate · Anthropic Claude · Amazon Nova Pro

Scale

• TransPerfect has over 10,000 employees.
  “TransPerfect has grown into an enterprise with over 10,000 employees in more than 140 cities on six continents.”article, stated
• TransPerfect operates in more than 140 cities on six continents.
  “Founded in 1992, TransPerfect has grown into an enterprise with over 10,000 employees in more than 140 cities on six continents.”article, stated
• AWS localization handles billions of words across digital assets.
  “The AWS team handles billions of words in multiple languages across digital assets. Given the growing demand for multilingual content by internationally minded businesses and new local cloud adoption journeys, the AWS team needs to support an ever-increasing load and a wider set of languages.”article, stated
• Bedrock LLM post-editing improved over 95% of suggested edits.
  “For post-editing, over 95% of all edits suggested by Amazon Bedrock LLMs showed markedly improved translation quality, leading to up to 50% overall cost savings for translations for Transperfect and freeing human linguists for higher-level tasks.”article, stated
• Post-editing produced up to 50% cost savings for translations.
  “For post-editing, over 95% of all edits suggested by Amazon Bedrock LLMs showed markedly improved translation quality, leading to up to 50% overall cost savings for translations for Transperfect and freeing human linguists for higher-level tasks.”article, stated
• LLM-powered transcreation produced up to 60% linguist productivity gains.
  “Using GlobalLink powered by Amazon Bedrock for transcreation, users are seeing linguist productivity gains of up to 60%.”article, stated
• Customers saw up to 40% workflow cost savings and up to 80% turnaround reduction.
  “customers in industries ranging from life sciences to finance to manufacturing have seen cost savings of up to 40% within their translation workflows and up to an 80% reduction in project turnaround times.”article, stated

Alternatives considered (roads not taken)

Reuse only translation memory and conventional machine translation

Rejected because: Unrecycled segments and post-editing still added days, cost, and turnaround time.

Fully manual human transcreation

Rejected because: Creative transcreation was expensive, slow, and resistant to earlier automation, so LLM candidate generation was used to reduce time and cost.

Fully automated transcreation selection

Rejected because: The workflow keeps human linguists selecting among generated candidates rather than letting the model publish creative copy unattended.

Patterns

human-in-the-loop-gate (quality)

Human linguists either perform lighter post-editing or choose among LLM-generated transcreation candidates.

guardrails-io-filtering (safety)

Bedrock Guardrails contextual grounding checks detect and filter hallucinated or inconsistent post-editing output.

prompt-versioning (prompting)

Post-editing prompts include style guides, approved translations, and errors to avoid so the model edits in the expected style.

safety-policy-enforcement (compliance)

Bedrock is chosen partly for data ownership, compliance, auditability, and sensitive-industry requirements.

model-routing (modeling)

Different translation modes route to translation memory, Amazon Translate, Bedrock LLM post-editing, or Claude/Nova transcreation depending on segment and content type.

Tradeoffs

Use LLM post-editing after translation memory and machine translation

Chose Bedrock LLM automatic post-editing with style guides and approved examples over only human post-editing after machine translation.

Rationale: The LLM step reduces cost and turnaround time while letting humans focus on higher-value edits where needed.

Keep human selection in creative transcreation

Chose multiple LLM candidate translations plus linguist choice over fully automated creative translation publication.

Rationale: Creative language needs cultural and emotional equivalence, so humans remain responsible for selecting the best adaptation.

Use Bedrock for secure enterprise LLM integration

Chose Amazon Bedrock inside GlobalLink over LLM services that share data with model providers or lack required auditability.

Rationale: Data ownership, FedRAMP authorization, compliance standards, logging, and guardrails matter for sensitive translation customers.

Gotchas (what bit them)

LLM hallucination is unacceptable in translation post-editing.

Root cause: Large language models can generate incorrect information, which breaks the requirement for precise translation.

Resolution: TransPerfect used Bedrock Guardrails contextual grounding checks to identify and filter inaccuracies.

Creative transcreation resisted earlier automation.

Root cause: Creative content depends on nuance, subtlety, humor, descriptiveness, and cultural references.

Resolution: The system asks Claude or Nova Pro for multiple candidate adaptations and leaves selection to a human linguist.

Post-editing can add days to localization timelines.

Root cause: Human linguists must finesse machine-translated output against style guides and glossaries.

Resolution: Bedrock LLMs perform automatic post-editing so linguists do lighter review or no-human-touch workflows can improve quality.

Evaluation

The case reports operational quality and efficiency outcomes for automatic post-editing, transcreation, and translation workflows, including quality improvement, productivity gains, cost savings, and turnaround reductions.

Metrics: over 95% of Bedrock-suggested edits improved translation quality · up to 50% translation cost savings for TransPerfect · up to 60% linguist productivity gains for transcreation · up to 40% customer workflow cost savings · up to 80% project turnaround-time reduction

Outcomes

• Automatic post-editing improved most suggested edits and reduced translation costs.
• Transcreation workflows saw up to 60% linguist productivity gains.
• Customers saw cost savings and turnaround reductions in translation workflows.
• No-human-touch machine translation workflows received a quality boost.

The deployment targets enterprise localization at billions-of-words scale across AWS assets and TransPerfect customer industries including life sciences, finance, and manufacturing.