Predictive Translation Quality Scoring for AI Localization Pipelines
MTQE (Machine Translation Quality Estimation) in Custom.MT helps localization teams evaluate MT output at scale before human review. It scores every translated segment on a 0–100 range, allowing your workflow to automatically accept high-quality segments and focus human effort only where it’s needed.
Result
Less manual review, lower costs, higher consistency, and faster delivery across all languages.
Traditional quality verification requires linguists to open, read, and judge every segment manually — a process that doesn’t grow.
With QE, your AI platform automatically answers:
“Is this segment good enough to publish?”
Control budgets by limiting human editing to low-scoring segments
Reduce manual review by skipping high-quality MT output
Create expandable, GenAI-augmented localization workflows
Improve linguistic consistency across vendors, teams, and markets
Add QE as a workflow step
Appears alongside Machine Translation, Post-Editing, and Terminology.
Choose your provider
Custom.MT supports multiple QE engines, including:
✔ TAUS – Available Now!
✔ ModernMT – Coming Soon
✔ OpenAI LLM QE – Coming Soon
✔ Claude LLM QE – Coming Soon
Set your quality threshold
Example: 85/100 as the minimum acceptable quality.
70–80 → Suitable for user-generated content, FAQs, and support articles
90–95 → For high-visibility content requiring near-perfect quality
You can adjust thresholds per client, language pair, or content type.
During processing
Each segment receives a 0–100 score.
≥ threshold → automatically approved (locked/confirmed in your CAT tool)
< threshold → sent to Automatic Post-Editing or human review
Confirm segments to mark them as finished while still allowing your team the freedom to make a quick, optional edit.
Lock segments to freeze the content entirely so that no human time or budget is spent on those segments.
Real-time Translation Quality Estimation (TQE) Scoring in your CAT tool
As you navigate your document, the Custom.MT pane at the bottom of your screen provides instant “Traffic Light” guidance for the active segment.
Combine MTQE + APE
Use the QE score to act as a smart filter for your Automatic Post-Editing engine. This ensures you only use AI rewriting resources where they are actually needed.
Evaluate quality threshold over time
Update your quality threshold based on:
Quality Estimation in Custom.MT is flexible enough to support different levels of automation — from fast, cost-efficient review to enterprise-grade quality assurance. Below are two recommended workflows depending on your content complexity, risk level, and required turnaround.
(Recommended for Most Teams)
MT → QE → APE → Human (only if needed)
MT generates the initial output.
QE (first pass) scores every segment and filters out high-quality results.
High-score segments → auto-approved
Low-score segments → routed for APE or human editors
APE (optional) improves only the flagged segments.
Human review happens only if required by content type.
Best for:
(For High-Risk or Regulated Content)
MT → QE → APE → QE → Human (only if needed)
MT produces the initial translation.
QE (first pass) identifies low-quality segments.
Automatic Post-Editing (APE) automatically refines only the low-score segments using LLM-based post-editing.
QE (second pass) re-scores the improved output.
Segments now above the threshold → auto-approved
Segments still below the threshold → sent to linguists
Human editors work only on the most problematic parts.
Best for:
○ Edit only the segments that need attention
○ Faster turnaround
○ Clear segmentation of "safe to publish" vs. "requires review"
○ Predictable quality
○ Lower vendor costs
○ Fewer review loops
○ Faster time-to-market
○ Risk mitigation
○ Growable quality control
○ Consistent standards across large content volumes
○ Integrates into any GenAI-driven localization strategy
Custom.MT supports several scoring engines, each with its strengths:
| Model | Description | Best Use Case |
|---|---|---|
ModernMT | Stable, fast scoring | General Workflow |
Quality Estimation, often called MTQE in the translation world, is an AI-powered process that predicts the quality of a machine-translated segment instantly without needing a human to check it. It assigns a score to each sentence to help teams decide which parts are safe to use immediately and which require a professional editor. This technology essentially acts as an automated “first look” that saves time by filtering out high-quality translations from the risky ones.
While MTQE (Machine Translation Quality Estimation) focuses on predicting accuracy during the translation process, AI LQA (Language Quality Assurance) is an automated “audit” that evaluates the final output against specific style guides and error categories. Think of MTQE as an instant health check for a translation, whereas AI LQA is a more detailed post-exam that mimics a human editor’s feedback. Essentially, MTQE decides if a translation can pass through the workflow, while AI LQA explains exactly why a segment might be failing.
No, but it reduces their workload. Humans focus only on segments below your quality threshold.
According to our benchmark, there is no single “best” provider, as performance varies significantly by language pair and business goals. General-purpose models like OpenAI and Claude often achieve higher raw scores across many languages, but specialized providers like ModernMT, TAUS, and Widn.AI are often easier to integrate into professional localization workflows.
The benchmark reveals that a “one-size-fits-all” provider doesn’t exist, so the most effective strategy is to test multiple systems against your own language pairs and content types. You can use the Custom.MT threshold optimization tool to get a personalized summary of which models perform best for your specific needs.
Right now, QE works on Trados, but we are going to expand across memoQ, XTM, Smartcat, and other connected tools.
Yes, you can auto-refine low-quality segments using APE prompts.
The optimal QE workflow depends on your content type and quality goals, but the most common setup is:
MT → QE → Post-editing (only for low-score segments)
This flow ensures that high-quality machine-translated content is approved automatically, while linguists focus only on segments that fall below the threshold.
It depends on your quality expectations:
You can adjust thresholds per client, language pair, or content type.
QE works well with both, but: