Global enterprises often struggle to maintain a unified brand voice when translating millions of words across dozens of markets. Terminology drift, which is the inconsistent use of product names, technical terms, or brand slogans, creates customer confusion and erodes trust. When a single technical component is translated three different ways across a user manual, customer support tickets rise and brand authority drops. To solve this, localization teams use data-driven metrics to track how closely translations follow approved glossaries, ensuring that speed never comes at the expense of precision. This quantitative approach guarantees that the core identity of the business remains intact, regardless of the language or the scale of the project.
Key takeaways
- EPT is the primary engine for measuring terminology quality, providing a clear numerical score of errors per thousand words.
- Time to Edit (TTE) decreases significantly when terminology adherence is high, as editors spend less time correcting repetitive glossary violations.
- Human-AI symbiosis is essential to distinguish between genuine errors and valid linguistic exceptions that automated tools might misflag.
- Glossaries are strategic data assets that improve the performance of purpose-built models like Lara when refined through adherence reports.
What terminology adherence actually tracks
Terminology adherence measures the frequency with which a translation aligns with a pre-approved list of terms, known as a glossary or termbase. It tracks whether the specific nomenclature that defines your brand, such as product names, technical specifications, and industry-specific jargon, is used correctly across every language pair. This process ensures that your “customer portal” in English does not become a “user gateway” in German and a “client hub” in French. Maintaining this lexical discipline is foundational to building semantic consistency in any global market.
Modern terminology tracking goes beyond simple string matching. It evaluates semantic accuracy within the context of the entire document. Advanced models like Lara, Translated’s purpose-built LLM, are designed to understand these nuances. Unlike generic models, Lara recognizes when a specific brand term must be used and when a synonym might actually hurt the user experience. By focusing on these core entities, enterprises can maintain a consistent identity while scaling their global reach. This level of understanding prevents the dilution of technical concepts and guarantees that the target audience receives the exact intended message.
How automated checks catch glossary violations
Automated quality assurance tools perform the heavy lifting in terminology management. These tools integrate directly with TranslationOS, the centralized hub where your global brand assets are stored. During the translation process, the system automatically cross-references the source text with the approved glossary. If the source contains a term that the translator failed to use correctly in the target language, the system flags a violation instantly. This immediate feedback loop prevents errors from traveling down the production pipeline.
These violations are then quantified using the Errors Per Thousand (EPT) metric. To calculate the EPT for terminology, you divide the total number of glossary errors by the total word count of the document and multiply the result by 1,000. In many frameworks, violations are also categorized by severity to properly weight the EPT score. A critical violation, such as using a forbidden competitor name, carries more weight than a minor capitalization inconsistency. For example, a 10,000-word manual with a weighted sum of 5 for terminology errors results in an EPT of 0.5. This objective score allows localization managers to set clear quality thresholds and identify projects that require immediate human intervention before they reach the final user. Establishing a strict EPT baseline gives organizations a mathematical foundation for measuring translation vendor performance and overall linguistic health.
Where legitimate exceptions get miscounted as errors
While automated tools are fast, they lack the linguistic intuition of a human expert. This often leads to false positives, where a tool flags an error that is actually a correct translation. In many languages, glossary terms must change their endings to fit the grammatical case or tense of the sentence. A strict automated check might flag a German pluralization of a product name as an error simply because it does not match the singular entry in the glossary. Languages with complex morphology require contextual interpretation that basic matching algorithms simply cannot provide.
Polysemy, where one word has multiple meanings, is another common cause for miscounting. If a glossary term like “drive” is intended for automotive context but is used in a section about human motivation, the tool may still demand the automotive translation. This is why human-AI symbiosis is essential. Professional reviewers must go through the QA report to whitelist these valid exceptions. This ensures that the final EPT score reflects true quality issues rather than the rigid limitations of software. Human experts provide the necessary contextual oversight to maintain the accuracy and integrity of the localization metrics, proving that technology alone cannot replace professional linguistic judgment.
Why this metric matters more as content volume grows
As content volume increases, the cost of terminology inconsistency compounds. In a small project, fixing a few terms is a minor task. However, for an enterprise localizing millions of words a month, a high rate of terminology errors can paralyze the production pipeline. This is where Time to Edit (TTE) becomes the critical KPI. When machine translation outputs consistently align with approved glossaries, the human editor spends less time on repetitive manual fixes and more time on high-level style and nuance. Efficient workflows rely heavily on getting the core vocabulary right on the first pass.
Lowering TTE directly correlates with a higher ROI for your localization program. By maintaining high terminology adherence, companies can push content to market faster without increasing their quality assurance budgets. When building professional localization workflows, this scalability is essential for industries like healthcare or finance, where a single incorrect term can lead to regulatory non-compliance or legal liability. In these high-stakes environments, the EPT score for terminology is a definitive metric for risk management, ensuring that rapid content delivery never compromises global safety standards.
Using adherence data to improve the glossary itself
The value of measuring terminology adherence extends beyond a simple binary report. These reports provide a wealth of data that can be used to refine your brand assets. If a specific term consistently triggers false positives or is repeatedly avoided by human translators, it may be a sign that the glossary entry itself is outdated or poorly chosen. By analyzing these patterns, you can optimize your terminology database to better reflect the evolving way your customers talk about your products. A proactive review process turns static glossaries into dynamic, responsive linguistic assets.
This data-centric approach is the foundation for improving AI translation models. When you feed high-quality, corrected data back into your ecosystem, purpose-built models like Lara become more accurate. They learn to handle specific brand terminology with greater confidence, further reducing the need for human intervention over time. Deploy adherence data to do more than track quality; actively build a more intelligent and efficient localization engine for your future growth. As the system absorbs these corrections, the overall localization process becomes inherently smarter and more aligned with your global strategy.
Frequently asked questions
These common questions explore the technical and operational aspects of measuring terminology quality in professional localization workflows.
What is a good EPT score for terminology?
For most enterprise marketing and technical content, a terminology EPT of 1.0 or lower is considered excellent. In highly regulated industries like legal or pharmaceutical translation, the threshold may be even stricter, requiring a score of 0.5 or lower to ensure total compliance and safety.
How does Lara handle terminology better than generic LLMs?
Lara is specifically fine-tuned for translation tasks and is designed to respect full-document context. While generic models may pick the most common synonym, Lara understands how brand entities relate to the surrounding text, allowing it to apply glossary terms more accurately even in complex sentences.
Can automated checks replace human QA?
No. Automated tools are excellent at catching literal mismatches, but they cannot evaluate the nuance of meaning or handle grammatical variations effectively. Human experts are necessary to validate exceptions and ensure that the final EPT score accurately reflects linguistic quality.
How often should glossaries be updated?
Glossaries should be treated as living documents. We recommend a formal review at least once a quarter or whenever a new product is launched. Adherence reports from TranslationOS can help you identify terms that are no longer performing well and need to be refined or replaced.
