In order to empower you to be on top of the quality of translations across different language pairs and content types, Unbabel designed the Quality Reports feature.
Access the Quality Reports app by heading to Reports > Quality in the navigation panel to the left.
Quality reporting allows you to check the quality performance of your translations, at any time, by applying state of the art Quality Estimation to determine quality scores and organizing them with the Customer Utility Analysis (CUA) framework.
CUA consists of five levels of quality, each being determined by a range of MQM scores. MQM (Multidimensional Quality Metric) is a standard quality framework that emerged from collaboration between academia and industry, designed to assess translation quality. For each different channel (chat, email, FAQ), we break down quality into predetermined buckets in which a certain percentage of your completed translations will fall in.
The report will include:
- An overview of quality across translations
- Total active language pairs (LPs) annotated
- Total translations made within the past 6 months
- Breakdown of quality per active LP
- In case the quality of any LP does not meet Unbabel's threshold (i.e over 10% in the red), the status of actions being taken to improve quality to meet the acceptance criteria
The five quality buckets are:
Best - Premium quality (MQM of >=98)
The translation is exceptional and displays outstanding fluency. Any potential errors are negligible and have minimal impact on the overall meaning and success of communication.
Excellent - High level of quality (MQM in the range 94 ≤ x < 98)
The translation is fluent and accurately conveys the intended meaning. Minor style or fluency issues may still occur, but communication is effective and successful.
Good - Good level of quality (MQM in the range 80 ≤ x < 94)
The translation is easily comprehensible and accurately conveys the intended message. Although there may be occasional grammatical errors, the overall quality is acceptable.
Moderate - Below target quality, area of attention (MQM in the range 60 ≤ x < 80)
This level requires attention, as errors may impact the clarity and understandability of the translations.
Weak - Poor quality (MQM of <60)
The translation requires review as errors will likely impact the comprehension of the message's core concepts.
How do we measure quality?
Quality score in the Portal is obtained from the score of all your requested translations in real-time. Unbabel is able to leverage large amounts of data by using it's state-of-the art Quality Estimation AI technology. Read more here.
Understanding the report
When entering the report page, the top chart on the board will include data pertaining to the sum of all your subscribed language pairs. Information about the total of translations, their percentage and quality score according to the CUA will be displayed here, giving you an overall idea of how the operation is performing. The report is built based on the quality from all translations of the past six months.
Below the Quality Overview chart, you can find the same information broken down by language pair. Languages are sorted by their relative volumes. Clicking on a language will display the CUA chart for the LP.
Having the data broken down by language pair will allow you to detect languages where the quality may require more attention - this can either require actions from your side, such as elaborating/reviewing more language resources to send us, or from Unbabel's. If the quality drops below a certain threshold, we will automatically detect it and our Lang Ops department will program the necessary corrective actions.
Click on "Learn about Translation Quality" to find out more about CUA analysis. If you want to know more about how we measure quality, you can click on "Learn about Quality Estimation".