David Koot

Hi. My name is David Koot. I work at the intersection of language and data: measuring, modelling and improving the quality of machine and human translation. Over the years that has gone by many names (applied scientist, data scientist, localization consultant, language engineer, trainer of language models), so on this page I would rather describe what I actually did, and what I find interesting about it.

I am working for TAUS for a long time now. TAUS was, when I started, a think-tank for the localization industry: a small company in the Netherlands, but very influential in the localization and translation industry. I already had long experience in localization jobs by that time, so I could tell something about translation quality. I was asked to bring the quality framework called DQF to the next level. It was, back then, merged with the larger framework called MQM, a set of rules and ideas on how reviewers can estimate the quality of a translation, by assigning types of errors to it and estimating the severity of those errors. Then some calculation is applied, and you get a score. I wanted to put those ideas into practice by creating, with the TAUS team, the instruments and tools to apply and see the results of those quality estimations, and at the same time working on the standardization of that framework, which later became the most well-known quality standard in translation: MQM.

With my team we created a new unit that could fit the complete lifecycle of translation projects, including review phases and correction steps, so that we could track productivity and quality metrics for translation projects from start to finish, making all projects of different kinds and shapes comparable. We tracked projects from a number of big companies, and published insight into where the localization industry was heading. That way we kept a finger on the pulse of the industry.

Business Intelligence bulletins

I had a lot of fun writing those insights in quarterly bulletins. Here are a few:

• A BI Bulletin about how productivity is affected by the quality expectation of translation (light editing or high quality translation).
• A BI Bulletin about how productivity is affected by the use of tools to translate: machine translation vs translation memory vs translation from scratch. Which one works faster, and can we see how fuzzy translation memory holds up against machine translation?
• A BI Bulletin that makes it clear that translation memories could be very efficient for post-editing translation, especially for high fuzzy matches.
• A BI Bulletin that compares productivity across different languages.
• A BI Bulletin about reviewing projects, and error types that occur on those projects. A very rare example of this type of research that looks at which error types are the most common in real-life translation projects.
• A BI Bulletin about some properties of translations that affect productivity. Among others, it introduces the concept of 'prose ratio', a metric for typical natural language, versus the type of translations that are full of numbers, tags, and symbols.

Where this leads

The common thread in all of this is data quality: if you can measure how good language data is, you can improve it, and you can build better models on top of it. More recently my work has moved further into machine learning, building models that estimate translation quality automatically. That is the direction I continue in: hands-on applied machine learning and NLP, where experimentation meets real problems, and where knowing the language side of the field is an advantage rather than the whole story.