Machine Translation Quality Estimation: Memsource's Latest AI-powered Feature

Remove the guesswork from machine translation with our latest feature powered by artificial intelligence.

We are excited to announce the release of our second feature powered by artificial intelligence, Machine Translation Quality Estimation (MTQE). Using MTQE, you can see MT quality scores before any post-editing is done, removing the guesswork from MT and improving post-editing efficiency.

The feature can be used with over 70 language pairs and is available for all 25+ MT engines supported in Memsource and in all Memsource editions. To use the feature you will need to purchase characters.

Initial testing indicates that for the supported language pairs, MTQE can provide quality scores of between 85%-100% in up to 14% of segments that are machine translated. In practical terms, this could mean up to 7% of savings on post-editing costs.

“Given the success of our first AI-powered feature, we wanted to keep up the momentum by launching another AI feature within the same year,” said David Canek, Memsource CEO. “MTQE builds on our AI-powered Non-translatables feature, with the aim of making post-editing easier and more efficient for everyone. Together they can equate to some significant cost reductions and increased productivity for our customers.”

MT Post-editing: The Problem

Machine translation quality continues to improve and linguists are increasingly being asked to edit, or post-edit, machine translation output. But machine translation quality still varies considerably depending on various factors, including content type and language pairs. Linguists have to sift their way through a lot of low-quality machine translation, occasionally coming across something that only requires a bit of editing or none at all. This wastes a lot of time.

We know that linguists are saving time with quality scores for translation memory matches so why not provide the same for machine translation. This is where AI, and our AI team, came in.

How MTQE was Developed

Our AI team looked into the historical translations where MT post-editing was used and based on the required post-editing effort, assigned the MT match a quality score. Our deep neural network was then trained with the historical data and thus learned how to accurately assign scores to a suggested MT output with any given source text.

Benefits of MTQE

Review quality of MT output in Memsource: The instant MT quality scores allow linguists to factor MT quality into the post-editing process and better manage low-quality machine translation which can increase post-editing efficiency.

Produce faster translations: With high-quality MT output, which receives a 100% score, post-editing may not be required at all.

Estimate post-editing effort and time: MTQE allows for a predictive analysis of MT performance with specific content. The resulting data can help assess post-editing effort and be used to create more competitive translation quotes.

Assess quality of MT engines: MTQE can be used to review the quality of MT engines for specific content so the engine with the highest quality scores can be selected.

MTQE in Memsource

Quality scores

The MTQE quality scores are available in the Memsource Web Editor and Desktop Editor. In MTQE Version 1, there are four scoring categories:

100%: Excellent machine translation quality, probably no need for post-editing

95%: Very good quality machine translation, possibly minor post-editing required

85%: Good match, but likely to require some post-editing

No score: When there is no score, this means MTQE cannot identify the quality so the output needs to be checked by a linguist.


Running an analysis with MTQE enabled will calculate the number of MT matches (think of these in the same way as translation memory matches) with scores of 85%, 95%, or 100% in the file. By expanding the 100%, 99-95% and 94-85% sections of the analysis table, users can see a breakdown of which matches are returned from translation memory (TM), which are from non-translatables (NT), and which are MT matches (outlined in red).


Machine translation, along with the quality score, is provided in two ways.

First, in real-time, segment by segment within the Memsource Web Editor and Desktop Editor (shown above).

Second, via the Pre-translation feature which allows batch pre-translation with a number of options. In the pre-translation settings, there is now the option to confirm 100% machine translation matches right away (as shown below).