Featured Translator: Carola F. Berger, PhD

  • Post author:
  • Post comments:0 Comments
  • Reading time:7 mins read
  • Post last modified:

Meet Carola F. Berger, an ATA-certified patent and technical translator and post-editor. Carola specializes in translations and post-editing between German and English. With extensive expertise in Machine Translation Post-Editing (MTPE) and neural machine translation, Carola has significantly contributed to the field. In 2022, she co-authored a presentation on quality metrics for the MT Summit, hosted by the American Machine Translation Association (AMTA). Additionally, Carola has delivered numerous presentations on artificial intelligence, machine translation, and post-editing for the ATA and other translation organizations.

Q: What initially drew you to translation, and how has your background in physics influenced your choice?

A: My second career as a translator happened purely by accident. I had some unexpected expenses that my tiny post-doctoral salary didn’t cover, so I was looking for ways to supplement my income. Somebody was looking for a translator from English into my native German, and the rest, as they say, is history. My extensive technical and science background naturally led me to my current focus on patents and texts in the fields of science and technology.

Q: Developing your own MT engine is a significant achievement. Can you share some insights into the challenges you faced during this process and how you overcame them?

A: I should mention right here that I failed miserably. That is, the neural MT engine ran just fine and training progressed as expected. But the output was essentially useless for my daily work, because as it turns out, to train any kind of neural engine you need a huge amount of data. My translation memory with over a dozen years’ worth of patent translations was by far not sufficient to train such an engine. However, the process of building and training the engine was very enlightening, because I gained a lot of valuable insights into how these engines really work, and what they can and cannot do.


Q: In your co-authored presentation at the MT Summit 2022, you discussed quality metrics for machine translation. Could you explain some of these metrics and why they are critical for evaluating MT output?

A: People have been trying to automate quality estimations of human and machine translation for a long time with various quality metrics. Sadly, some of these metrics are very outdated and not really applicable to neural MT output, because they are based on statistical analysis and comparisons to a reference translation. While this allows these metrics to be automated, it makes them less useful for gauging neural MT output and human translations. Not surprisingly, that’s what we found in our study, namely, statistical QE metrics requiring a reference translation strongly depend on which reference is used.

Now, this result is not earth-shattering, but in another aspect of our study, we were trying to see how these metrics compare to manual, human evaluations of translation quality. One of the results was that non-statistical QE metrics such as COMET and human evaluations align quite well if the translation to be evaluated is of “average” quality. However, the automated and human evaluations diverge quite a bit for translations that are either excellent or absolutely horrible. That said, our study was fairly small in scope, but could certainly serve as a starting point for further studies.

Q: As a seasoned presenter on topics related to AI and translation, what do you see as the most exciting or promising AI developments affecting the translation industry today?

A: There is a potentially promising AI development that I can envision, which, to my knowledge, does not exist as of yet, at least not as a publicly available application:

Right now, many LSPs, in an effort to be as cost-efficient as possible, often use platforms where translators register with their expertise, language combinations, and rates. Then LSPs post translation projects on these platforms, and the translator selection process is often less than ideal, especially for large LSPs with a large pool of translators, where project managers cannot possibly be familiar with the entire pool of translators. But here’s where AI could come in. AI engines are very good at making sense of a large amount of data and detecting patterns therein. LSPs collect a large amount of data on their translators. I can envision an AI engine which analyzes the collected data and then proposes a small pool of translator candidates to the project manager who are most qualified for the specific project, in terms of expertise, experience, credentials, and—yes, rates. Sounds simple enough, and perhaps a proprietary version of this is used internally by some of the very large LSPs, but the standard platforms that are used by a lot of LSPs without their own proprietary platform do not have any such feature.

Q: What motivates you to teach and share your knowledge with others in the industry?

A: I guess I can’t deny that I come from at least 4 generations of teachers. My great-grandfather was a teacher. My grandparents were middle school principals. My parents were high school teachers, as were my aunt and uncle before retirement. My sister is a teacher, as are all my cousins. I’m the black sheep of the family, a scientist turned translator. But teaching seems to be in my genes.

Q: What advice would you give to beginner translators on how to succeed?

Two things:

1.) Learn how to program computers. View coding as another language. The translators of the future will need to know their way around LLMs, generative AI, neural MT, CAT-Tools, terminology management tools, regular expressions, quality estimation tools, and a host of other technologies with funky acronyms that haven’t been invented yet. Don’t be afraid to try things out. The worst that can happen is that you crash your computer. And the computer doesn’t care, you reboot it and continue where you left off.

2.) Never stop learning. Language continuously evolves. So does technology. But despite all the technological advances and claims of the robots allegedly reaching human parity, I firmly believe that there will always be a need for human ingenuity until we reach the “singularity,” when machines can really think. As of right now, they can’t, even though they sometimes mimic humans well enough to fool us in believing that they can.

Carola’s insights and experiences highlight the dynamic and evolving nature of the translation industry, emphasizing the importance of continuous learning and adaptation. We are grateful to Carola for sharing her expertise and look forward to her presentation at our 9th International Virtual Conference. Join us in June to learn more on Machine Translation Post-Editing – Facts, not Fiction! 

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.