TL;DR
Machine Translation Post-Editing (MTPE) bridges the gap between pure AI speed and human-quality perfection by having professional linguists refine machine-generated text. Depending on the content’s purpose, enterprises utilize either Light Post-Editing (LPE) for quick, internal accuracy, or Full Post-Editing (FPE) for polished, customer-facing material. To eliminate subjectivity and measure ROI, localization teams track technical editing effort through metrics like Error Per Thousand (EPT) and temporal effort via Time to Edit (TTE). Ultimately, modern AI technologies, such as Translated’s context-aware model Lara and the TranslationOS platform, use this human editing data as feedback, turning MTPE into a continuous loop that steadily improves future automated translations
What is machine translation post-editing (MTPE)?
Machine Translation Post-Editing (MTPE) is the process of having a human linguist review and refine machine-generated translations. Its goal is to correct errors, improve fluency, preserve meaning, and ensure the translation fits the required context, while reducing the time and cost of translating from scratch.
For enterprises, this makes MTPE a production workflow that can also generate measurement data for evaluating MT quality and human effort. It helps teams answer practical questions: Is the machine translation good enough? How much editing does it require? Which types of content need human intervention? And how can AI systems improve over time?
To answer these questions, organizations need to connect MTPE workflows with MTPE metrics.
Where MTPE fits within translation workflows
Translation workflows sit on a spectrum. At one end, raw MT offers speed and scale but limited quality control. At the other hand, fully human translation provides maximum linguistic oversight but requires more time and resources. MTPE sits between the two: AI produces the first draft, and human linguists apply the needed level of review. This can be:
- Light Post-Editing (LPE): A reviwer makes only the changes needed for the text to be understandable and accurate. The goal is usability, not stylistic perfection. This makes LPE suitable for internal documents, user-generated content (often also translated with pure MT), or support material where speed matters more than polish.
- Full Post-Editing (FPE): A reviewer edits the MT output until it meets publishable quality standards, including meaning, grammar, terminology, tone, and cultural nuance. This is better suited for customer-facing or high-visibility content such as websites, product interfaces, marketing, and external documentation.
When deciding on translation quality, enterprises need to determine the right level of MTPE for each use case, depending on the required balance of speed, scalability, and accuracy. Under-editing can lead to inaccurate or ineffective translations, while over-editing can waste resources by refining content beyond what its purpose requires, undermining the efficiency gains of machine translation.
Why MTPE measurement matters
Without metrics, MTPE can become subjective. One linguist may make minimal edits, while another may rewrite the same text extensively. MTPE metrics create a shared way to evaluate:
- how good the raw MT output is,
- how much human effort is required,
- which errors occur most often,
- how long editing takes,
- whether the final translation meets quality standards,
- and whether MTPE is more efficient than translating from scratch.
This makes MTPE measurement useful not only for quality control, but also for workflow planning, cost estimation, vendor evaluation, and AI model improvement.
The MTPE process: Step by Step
MTPE works best when it follows a structured process. This makes post-editing repeatable, measurable, and scalable across languages, teams, and content types.
Step 1: Define the MTPE workflow
Project managers or localization managers predefine whether to use LPE, FPE, or fully human translation, depending on the project scope and content type. In modern localization pipelines, AI tools such as quality estimators can also analyze the machine translation output before editing begins. Based on predefined quality thresholds, the content is then translated following the most appropriate workflow: LPE, FPE, or fully human translation.
Step 2: Correct errors
The editor fixes mistranslations, omissions, grammar and spelling errors, punctuation issues, awkward phrasing, and terminology inconsistencies. Edit-based metrics can help estimate technical effort by tracking changes such as insertions, deletions, substitutions, and shifts.
Step 3: Apply style and consistency
The editor checks tone, terminology, style-guide compliance, and locale conventions. For enterprise content, this step helps distinguish a translation that is merely correct from one that is brand-consistent and ready for publication.
Step 4: Use localization tools and QA checks
Relying on translation memories, glossaries, style guides, and quality-assurance checks, editors can reuse approved translations, apply terminology consistently, and avoid correcting the same issues repeatedly.
Step 5: Measure MTPE
Finally, teams review editing data, quality results, time spent, and recurring errors. These insights help improve MT engines, glossaries, style guides, and routing decisions, including when to use raw MT, MTPE, or human translation.
How to measure MTPE: Key metrics
MTPE metrics evaluate two related things: translation quality and human editing effort. A practical way to organize them is to separate effort into two dimensions:
- Technical effort measures the number and type of edits made by the post-editor, such as insertions, deletions, substitutions, and word shifts.
- Temporal effort measures how long editing takes, often as seconds per word or overall time to edit.
Together, these dimensions help teams understand MTPE quality and efficiency. Two useful metrics for this are Error Per Thousand (EPT) and Time to Edit (TTE).
Error Per Thousand
Error Per Thousand (EPT) is a standardized translation-industry metric used to measure linguistic accuracy by calculating how many errors are identified for every 1,000 reviewed words during a Quality Assurance (QA) review. During this process, reviewers evaluate the translated text, classify the errors, and usually assign them a severity level through error points. The EPT score is calculated using the following formula:
EPT = (total error points × 1000) / reviewed words
In the translation industry, tools like MateCAT , Translated’s open-source CAT tool, support QA workflows by using EPT as an objective benchmark to monitor translation quality and identify recurring or systemic issues.
In those tools, project managers and reviewers can assign error points through a quality framework made up of issue categories and severity levels. These categories and severities can be modified or expanded, allowing teams to customize the quality framework and adjust EPT thresholds. This makes it possible to assign different weights to different types of issues depending on their impact.
Typically, standard issue categories include:
- Style: readability, consistency, and tone
- Tag issues: mismatches and whitespace problems
- Translation errors: mistranslations, additions, and omissions
- Terminology and translation consistency
- Language quality
Time to Edit
Time to Edit, or TTE, is a metric introduced by Translated. It measures how long a professional linguist spends editing machine-translated content until it reaches the required quality level. Although TTE is mainly used to evaluate machine translation quality, it can also be used to evaluate the temporal effort required for a translation. For enterprises, this makes TTE valuable because it connects translation quality to operational impact: lower TTE means less human editing time, faster turnaround, lower localization costs, and a clearer way to compare the real productivity gains of MT systems.
How Translated’s AI systems support post-editing workflows
Translated treats post-editing effort as a key indicator of translation quality. Among metrics, the company considers metrics EPT and TTE. In particular TTE is valued in relation to Tranlated’s long-term goal of reaching “translation singularity”: the moment when AI translation reaches a level where top professional translators spend the same amount of time revising AI-generated translations as they would revising work produced by another expert human translator.
To support this goal, Lara, Translated’s AI translation model, is designed to produce context-aware translations that help reduce repetitive correction work. Rather than treating each sentence in isolation, Lara uses broader document context to improve flow, terminology consistency, tone, and gender agreement, allowing linguists to focus more on nuance, cultural adaptation, and final quality decisions.
As an adaptive translation system, Lara can also use human feedback for improvement, enabling human-in-the-loop workflows. In this, post-editing becomes more than a correction step: AI produces the first draft, linguists refine and validate it, and the resulting feedback can help future translations better reflect enterprise terminology, tone, and quality expectations.
Additionally, TranslationOS, Translated’s AI service delivery platform, supports MTPE-style workflows by acting as a central environment for localization workflows, linguistic assets, and performance data. By bringing translation memories, glossaries, style guides, AI output, and human review into the same system, it helps teams evaluate both translation quality and production efficiency across large-scale projects. In this way, linguistic assets become active parts of the post-editing process rather than static reference materials.
Conclusion
MTPE measurement helps enterprises move beyond subjective quality judgments by connecting human editing effort, translation quality, and operational efficiency.
The value of MTPE measurement is that it connects linguistic quality with operational decisions. By tracking editing effort, error patterns, terminology issues, and production time, localization teams can choose the right workflow for each content type, improve translation resources, and identify where machine translation and AI deliver the greatest efficiency gains.
With adaptive systems such as Lara, platforms like TranslationOS, and tools like MateCAT, post-editing can also become part of a continuous improvement loop, where human expertise helps refine future translation output while preserving quality, consistency, and brand voice.
