Note: This article describes how this feature works in Lokalise Expert. The same feature is also available in Lokalise Vantage but the overall workflow differs. To learn more, see: Lokalise Vantage: Onboarding, Lokalise Vantage projects, Lokalise Vantage: Editor.
This feature is available from the Growth plan and above.
The AI-powered translation scoring feature helps you evaluate translation quality by identifying potential issues and assigning a score from 0 to 100 based on Multidimensional Quality Metrics (MQM).
High score (≥ 80): Few or no issues detected. The translation may require little or no human review.
Low score (< 80): More significant or multiple issues detected. Human review is recommended.
A high score doesn't guarantee a perfect translation. It simply means that AI scoring detected fewer or less severe issues.
You can request scores for individual translations in the editor or enable scoring for eligible Pro AI translation tasks. AI scoring is optional and disabled by default for tasks created in the Lokalise interface.
This helps you focus human review where it's needed most, rather than reviewing every translation or skipping reviews entirely.
What translation scoring offers?
With translation scoring, you can:
Focus human effort only where it matters, cutting review time and cost
Get instant, data-driven QA directly in the editor
Improve reviewer productivity with specific issue guidance
Confidently handle low-priority languages that were previously cost-prohibitive
Unlike generic scoring tools, Lokalise scoring is:
Integrated – built into the Lokalise editor and workflows
Context-aware – tailored to your project’s content and quality needs
Real-time – scores appear instantly per string
Cost-efficient – reduces unnecessary reviews and budget spend
Can we really trust an AI score to determine translation quality?
The goal of Translation Scoring isn’t to replace human reviewers: it's to help them focus on the content that truly needs attention. Instead of reviewing every string manually (slow and expensive), teams can:
Route only low-quality translations (<80) for human review
Skip or lightly edit high-scoring translations
Spend review time more strategically, without sacrificing quality
In practice, teams can skip or lightly edit high-scoring translations and only review lower ones, reducing post-editing effort by up to 80% while maintaining quality.
Why do you use MQM to evaluate translation quality?
We use MQM (Multidimensional Quality Metrics) because it’s the most widely recognized standard for evaluating translation quality with AI.
Unlike automatic metrics such as BLEU, COMET, or METEOR that compare output against a single "reference" translation, MQM evaluates the translation itself. It highlights errors, explains their severity, and produces interpretable scores that reflect real-world usability, not just a number.
How to view scores for translations
To see translation scores, open your Lokalise project editor and look for a target translation value. You’ll notice a small lens icon next to it:
This icon only appears for target translations, not source (base) values.
Click the lens to trigger scoring. AI will check the translation quality and show a score. Click the score to get more details. If the translation has issues, points will be deducted, and you’ll see a list of detected problems:
Scoring does not consume any extra AI words from your team’s quota.
Access AI scoring programmatically
You can retrieve AI scoring information programmatically using the Lokalise MCP server or REST API.
With the Lokalise MCP server, ask your AI agent to retrieve keys or translations and include their AI scoring information.
With the REST API, AI scoring information is available when listing keys, retrieving a key, listing translations, or retrieving a translation. Set
include_ai_scoring=1to include this information in the response. For key endpoints, also include translations in the response. When available, the results are returned in theai_scoringobject for each translation.
Scoring and AI tasks: Generating scores in bulk
You can generate scores for many translations at once. To do this, select one or more keys in the editor, then choose Create a task from the bulk actions menu.
Bulk actions menu
Bulk actions menu
In the task settings, select Automatic translation as the task type. To assess the quality of translations generated by AI, enable Trigger AI quality scoring with task. This option is disabled by default.
When enabled, Lokalise evaluates eligible AI-generated translations and assigns quality scores from 0 to 100. Scoring does not consume additional AI words from your quota.
AI scoring is available for Pro AI translation tasks when the task's source language matches the project's base language. Translations retrieved from 100% translation memory matches are not scored.
Scoring won’t apply to 100% matches from the translation memory. If a translation comes straight from your translation memory, it won’t be scored. Lokalise only scores translations generated by AI within tasks.
Using scoring in workflows
You can also set up scoring as part of your workflows. To do this, go to the Workflows page and create a new workflow that includes the Review task with AI scoring step. This step comes after the AI translation step, which may be preceded by the TM step.
When setting up the AI scoring step, you’ll see the following options:
View AI scoring step options
View AI scoring step options
You can set a quality score threshold. By default, it’s set to 80. If a translation score falls below this value, the translation is automatically added to a review task created during this step. Any translations that could not be scored for any reason will also be added to the task.
Keep in mind that scoring ignores translation memory entries with 100% matches. Lokalise only scores AI-generated translations within tasks.
Optionally, select Exclude reviewed translations. When enabled, translations already marked as Reviewed won’t appear in the reviewer’s queue.
You can also customize the review task: give it a name and description, set a due date, and assign team members to handle the translations flagged for review. For more details on review tasks, check the Tasks documentation.
How scoring works
Translation scoring is powered by the MQM (Multidimensional Quality Metrics) framework — the industry standard for evaluating translation quality. MQM breaks down issues into categories such as grammar, spelling, fluency, terminology, and meaning errors. Each issue type is assigned a penalty weight based on severity:
Critical (−75 points): Serious issues that make a translation unsuitable for use or could cause legal, safety, financial, or reputational harm.
Major (−25 points): Significant issues that change the meaning or make the translation difficult to understand.
Minor (−5 points): Noticeable issues that don't change the meaning or prevent understanding, such as minor grammar or phrasing problems.
The final score is calculated as:
100 – (total penalties)
Examples:
One major (-25) + one minor (-5) → Score: 70
Two criticals (-75 each) → Score: 0 (scores cannot go below 0)
Score thresholds:
100: Perfect translation, ready to publish.
80–99: Good quality, minor improvements possible but safe to release.
Below 80: Review strongly recommended.
In workflows, thresholds can be configured to auto-route low scores to post-editing, while high scores can be published with confidence.
Translation scoring vs. AI LQA
Both translation scoring and AI LQA use the MQM framework to evaluate translation quality, but they serve different purposes:
Translation scoring
Built directly into the editor with real-time scores for each string
Integrated into workflows, enabling automation (e.g., routing low scores to human review)
Provides a faster, more user-friendly experience for day-to-day translation work
AI LQA
Generates quality reports in batch format
Useful for large-scale assessments and audits
Offers less interactivity compared to scoring
Translation scoring is the natural evolution of AI LQA bringing the same evaluation logic into a more scalable, real-time, and workflow-integrated experience. Over time, it may fully take the place of AI LQA for most use cases.
Frequently asked questions
💡 Looking for more?
Find all answers in AI: Frequently asked questions!