Localisation 5 August 2026
The LISA QA Model: The First Standard That Scored Localisation Quality
The LISA QA model was the first to measure localisation quality with a standard score, and it is the ancestor of today’s error-based review. We examine its categories, its weights and its legacy.
Saying that a translation is “good” is easy; turning that judgement into a measurable score is hard. The first tool to put this measurement on a standard footing in the industry was the LISA QA model. Published by the LISA organisation, which no longer operates, this model was the first widely used standard to measure localisation quality by counting and weighting errors, and it laid the foundations of how review is understood today.
Contents
This article looks at how the LISA QA model worked, which error categories it covered, what the critical-major-minor weighting meant and why it still matters. The model itself has passed into history, but its logic lives on inside frameworks in use today, such as MQM-DQF. Understanding it is therefore the key to understanding current quality review.
What is the LISA QA model?
The LISA QA model was the industry’s first standard error model for localisation quality analysis. It was a database-driven, integrated and customisable application. It worked on an error basis: a predefined but open error typology covered linguistic, formatting and functionality errors. Each error type was assigned a penalty score, and the total was compared against a set threshold.
The model’s strength lay in its flexibility. For each project or translation type, evaluators could set a threshold for the number of minor, major and critical errors allowed. Quality was therefore treated as a context-dependent concept rather than an absolute one. A social media post and a legal text did not have to be reviewed with the same strictness.
Being database-driven mattered too. Every error found was recorded and stored together with its category and severity. These records did not merely decide whether a given text was accepted; they also became a learning resource for future projects. Seeing which error type recurred most often allowed a team to fix the weak points in its process. Review thus became more than a pass-or-fail gate: it became an instrument of continuous improvement.
Why was a model needed?
Quality is a blurred concept that everyone in translation studies defines differently. As House puts it, assessing the quality of a translation presupposes a theory of translation; different theories lead to different understandings of quality. The industry, however, needed a shared, repeatable and comparable measure. The LISA QA model filled that gap.
The model turned a subjective “I liked it, I did not like it” verdict into an objective process. The evaluator flagged errors while reading the text, and the system counted and weighted them. Two different reviewers could therefore arrive at closely comparable results for the same text. That repeatability made it possible for corporate clients to write quality into a contract.
The need grew as the industry gained scale and complexity. Assessing every version of a product released into dozens of languages at once against the same criterion was impossible without a shared framework. Where different reviewers worked in different countries with different expectations, everyone had to name the same error in the same way. By establishing this common language, the LISA QA model gathered scattered practice into a single framework.
Error categories: language, format, function
The typology of the LISA QA model covered three broad areas. Linguistic errors involved problems of translation and expression: the wrong term, inconsistent terminology, grammatical and stylistic flaws. Formatting errors concerned the appearance of the text: incorrect capitalisation, broken spacing, faulty number or currency formats. Functionality errors were technical problems that affected how the product worked.
This threefold division still echoes in today’s frameworks. Functionality errors are usually the most serious, because a broken link or corrupted code can block the user’s interaction with the product altogether. Linguistic errors, by contrast, may be noticed by the user without making the product unusable. The difference in severity shows why weighting is necessary.
Because the categories were open and customisable, the model could adapt to different kinds of localisation. In a software project, formatting and functionality errors came to the fore, while in a marketing text style and terminological consistency could weigh more heavily. Evaluators could choose which categories to focus on according to the needs of the project. This adaptability saved the model from being tied to a single content type.
Weighting: critical, major, minor
At the heart of the model was the scoring of errors according to their severity. The following penalties were assigned to each error type: a critical error 10 points, a major error 5 points, a minor error 1 point. The weight of an error mattered as much as the count. A hundred small errors and three critical ones could produce very different outcomes. The approach acknowledged that counting errors alone was not enough.
The idea behind weighting is that errors are not equal. A missing accent on a word and an instruction that misleads the user cannot be put in the same basket. Critical errors can make a product unfit for its purpose and may even have legal consequences, whereas minor errors may not even be noticed by the average user. The scores translate this reality into numerical terms and keep priorities in the right order.
This logic carried over into later models; only the scores changed. The table below compares the penalty scores of the LISA QA model with the default scores of today’s MQM metric. As can be seen, the principle is the same and only the scale differs.
| Severity level | LISA QA score | MQM default score |
|---|---|---|
| Critical | 10 | 100 |
| Major | 5 | 10 |
| Minor | 1 | 1 |
| Neutral | – | 0 |
The threshold value and the “unpublishable” verdict
The weighted error scores were added up and compared with a predetermined threshold. If the total exceeded the threshold, the localised text was deemed “unpublishable”. This threshold could vary with how critical the project or the content type was. Whether a text was accepted therefore rested on an explicit numerical rule rather than on a personal judgement.
The logic of the threshold also overlaps with the idea of “fit-for-purpose quality”. Absolute perfection is not the aim; the aim is a level of quality that stays below a given threshold and is considered acceptable. This pragmatic approach suited the reality of an industry working with limited budgets and limited time, and it still holds today.
Where the threshold is drawn depends on how critical the content is. For an instruction that could affect someone’s life, or for a text that could have legal consequences, the threshold is kept very low; for a short-lived promotional text it can be more relaxed. This flexibility is not arbitrary: it is justified by criteria such as content type, audience and the expected lifespan of the text. Documenting the threshold decision openly keeps the client and the team working to the same expectation.
The legacy LISA left behind
The LISA organisation closed, but the idea behind the model survived. The MQM-DQF framework, the most widely used standard today, was born from the merger of the dynamic quality framework developed by TAUS with the multidimensional quality metrics, and it inherited the error-based approach of the LISA QA model. A team that wants to adopt a model can follow these steps:
- Choose an error typology that suits the content type.
- Define a severity and a weight for each error type.
- Set an acceptable threshold for the project.
- Store the review results for future projects.
What sets modern frameworks apart is that they can be integrated into tools as plugins and that they offer a far more detailed list of issues. Even so, the basic steps follow the skeleton LISA built decades ago.
The limits of the model and holistic assessment
Error-based models are powerful, but they cannot measure everything. A text is more than the sum of its individual errors; its overall fluency, consistency and fitness for purpose matter too. Some assessments therefore use “holistic” criteria: the text is given a single overall score from 0 to 5. This method delivers a faster and more rounded judgement than counting errors alone, but it is less common in the industry.
Another limitation of the LISA QA model is its tendency to focus heavily on functional aspects and lightly on linguistic subtleties. As Dunne points out, because the functional side of software can be measured objectively, people tend to see localisation as a production process; linguistic quality, however, is often subjective. A good review has to balance numerical strictness against that subjective judgement. The model is a tool; it does not take the place of the assessor’s judgement.
The Türkiye context
Most translation agencies in Türkiye use the logic of the LISA QA model intuitively, even if they never mention it by name: find the errors, judge them by severity, then accept or reject against a threshold. As corporate clients turn towards standards such as ISO 17100, the need to convert this intuitive approach into a documented framework is growing.
When building a severity scale for Turkish, the particular features of the language have to be taken into account. In an agglutinative structure, a single wrong suffix can sometimes change the meaning completely; an error of that kind should be weighted far more heavily than a formatting flaw. Adapting threshold values to the content type keeps review consistent even in a small team.
What does this mean for you?
If you are buying translation, the presence of an error-based model gives you a concrete assurance: quality is assessed with a score rather than a feeling. On a project you can ask for the following:
- Find out which error typology your text will be reviewed against.
- Agree an acceptance threshold that suits your content type.
- Clarify how critical errors are defined.
We can settle on the right quality framework for your project together. You can browse our localisation services and see our pricing page for current terms. Our aim is to make quality measurable and open to discussion.
Frequently asked questions
Is the LISA QA model still in use?
The LISA organisation closed, but the logic of the model lives on. Error-based frameworks in use today, such as MQM-DQF, are direct heirs to the category and weighting approach of the LISA QA model.
What do critical, major and minor errors mean?
They are severity levels for errors. In the LISA QA model a critical error was weighted at 10 points, a major error at 5 and a minor error at 1. If the total exceeded the set threshold, the text was deemed unpublishable.
Which errors did the LISA QA model cover?
With an open and customisable typology, it covered linguistic, formatting and functionality errors. Evaluators could set their own thresholds for each project type.
What is the difference between the LISA QA model and MQM?
The basic logic is the same: find the errors, classify them, weight them. MQM-DQF offers a more detailed issue typology and different scores, and it can be integrated into tools as a plugin.
References
Jiménez-Crespo, M. A. (2024). Localization in Translation. Routledge (Chs. 3/4/5/9/12). · Lommel, A. (2018). Metrics for Translation Quality Assessment. Translation Quality Assessment, Springer.