Machine translation: how it works and fails
Understand how machines translate, judge translation quality and know when a human translator is needed
A taste of a lesson
I translated a Turkish text into English and back again, and it came back almost identical. So the English is correct?
Unfortunately that does not prove it. If the system makes a consistent mistake, such as choosing 'he' for the gender neutral Turkish 'o', it can undo the same mistake on the way back, so the round trip looks perfect. Back translation also hides lost nuance, tone and formality. Better checks: have a bilingual person review key passages, focus on numbers, names, negations and gendered words, and use a glossary for important terms. For anything legal or medical, use a professional translator. Quick exercise: find one sentence in your text where the English had to choose information the Turkish did not state.
Written by the teacher as an example. In your lesson the tutor answers your own questions, and like any AI it can be wrong.
What you will be able to do
- Describe how rule based, statistical and neural translation differ
- Distinguish fluency from adequacy and spot meaning errors in fluent output
- Explain what BLEU, chrF and learned metrics can and cannot show
- Use glossaries, context and post editing to improve results
- Decide when a professional human translator is essential
Lesson plan
- 1 From rules to neural networks Follow how machine translation evolved and what each approach got right. Start
- 2 Fluent but wrong Separate fluency from adequacy and find meaning errors in smooth output. Start
- 3 Gender, formality and culture Recognise choices a translation system makes that the source did not specify. Start
- 4 Measuring translation quality Use automatic metrics sensibly and combine them with human review. Start
- 5 Glossaries, context and post editing Set up a workflow that gives consistent, reviewed translations. Start
- 6 When a human translator is essential Identify content where machine translation alone is not acceptable. Start
Try asking
About this tutor
For anyone who uses or relies on machine translation: travellers, businesses, students, translators and developers. You will follow the move from rule based and statistical systems to neural encoder decoder models and general language models that translate, and learn why modern output is fluent but can still be wrong. The lessons separate fluency from adequacy, explain automatic metrics and their caveats, and cover recurring problems: gender and formality choices, idioms, inconsistent terminology and dropped negations. You will set up glossaries and a post editing workflow, and learn when a professional human translator is essential, such as for legal, medical or published material.
Reviews
4.3
3 ratingsSample
- Selin K.Sample
The back translation example with gender neutral Turkish was exactly my situation. I now get a bilingual colleague to check key paragraphs.
- Pierre L.Sample
As a translator I expected to dislike it, but it was fair to the profession. The post editing lesson matches how I actually work.
- Dong H.Sample
Clear explanation of metrics. I would have liked more on Korean specifically, but the general checks still apply.
About the teacher
Practical NLP: from tokens and embeddings to classification, translation and speech
9 tutors 470 lessons taught Sample
I teach natural language processing as a craft: turning messy text in many languages into something a model can use, and checking honestly whether the result works. I grew up switching between Arabic, French and English, and my work has been on text and speech systems that had to serve speakers of more than one language, so I notice quickly...
See Nadia's profile and tutorsMore like this
Other tutors on the same or nearby topics.