Ildikó Szemők’s Post

Why You Still Need a Human Translator (Part 7) AI is fast. But professionals add real value. Take this simple line: “Open your Twitter account.” AI translated it as: ➡️ “Open je Twitter-account.” Looks fine, right? But Huisman Translations delivered the correct, up-to-date version: ➡️ “Open je X-account.” That attention to detail makes all the difference. AI gets you close. A professional gets it right. Huisman Translations, human translations. #Translation #Localization #l10n #t9n #Translators #LanguageServices #EnglishDutch #PostEditing #HumanTranslations

AI and non-human translation, they're good for some things… but not for theatre! I was translating a theatre text the other day when 'Cállate' came up. On the surface, this seems simple to translate... at least that's what a computer would think, and mechanically chuck out 'Shut up' as a translation. But how to explain to a machine the context behind this moment in the play: 🕑 the time period; 🎩 the high social register being momentarily broken in this instance of frustration and power loss; 😡 the fact that the stem of the word is used by another character shortly after to highlight the animosity of the first and make the audience feel uncomfortable in doing so; 🏋♀️ not to mention 'shut up' not carrying the serious weight of this moment. Beyond this, there are also obvious breaches of copyright if you plug a literary text into a machine without adequate levels of privacy. Human translators bring an awareness of society, culture, voice, tone, register and, most importantly for theatre translation where the text is performed by and for humans, they know how to sound human. ⭐ I'm proud to say that I'm a theatre translator, and I'm proud to say that AI can't replace what I do. 🎉 Now that’s a good way to start the New Year! Happy New Year, everyone! ❓ What do you think? #Translator #Translation #Challenge #AI #IA #ArtificialIntelligence #InteligenciaArtificial #HumanTranslator Chartered Institute of Linguists Institute of Translation and Interpreting Out of the Wings

‘Just read more fiction,’ I said. Then 2025 happened Six years ago, a junior translator asked me for advice: “What can I do to improve my style? How do I stop sounding stiff, robotic?” Back then, my answer was honest and obvious (to me): “Read more fiction. In your target language.” That was the advice I truly believed in. But here we are — in a world where: * even clearly flawed localization is often accepted as “good enough”, especially when key stakeholders decide it’s “fine” * human translation has quietly become a premium choice, that many companies believe unnecessary  * many clients default to machine translation * and “sounding human” has become a blurry concept — increasingly defined by AI detectors If a junior translator asked me today — in 2025, on the edge of 2026 — what would I say? “Don’t worry too much, just learn to work with AI” (?) “Go teach AI to sound human — that’s where linguistic intuition actually matters now” (?) “Maybe even consider a field where AI won’t crawl in next year… a handyman, construction worker, housewife (ha-ha)” (?) “You’ll clean up machine translation and make it sound human. So keep reading. Human style still comes from humans” (?) What advice would you give a junior translator starting out today?

Do you know the difference between CAT (Computer Assisted Translation) and AI tools? They often get lumped together, but they aren’t the same. To start, CAT tools (Computer-Assisted Translation) don’t translate. They support human translators by storing past translations, managing terminology, and ensuring consistency across large or recurring projects. The benefit of CAT tools is that they improve accuracy and efficiency, while relying on human judgment and refinement. On the other hand, AI (or Machine Translation), generates translations. It can be helpful in producing a very rough draft, but it doesn’t understand nuance, cultural expectations, legal implications, or technical meaning. When used in regulated or high-risk industries, there are real concerns with both data protection and accuracy. Machine Translation requires a professional linguist to review every line. In fields like legal, government, healthcare, and technical manufacturing, it’s the human expertise, not the tool, that leads to accuracy, clarity, and compliance. If you’re choosing a translation partner, ask not just what technology they use, but how their linguists work with the tools.

Why You Still Need a Human Translator (Part 9) AI is fast. But professionals add real value. Take this simple term: “soccer uniforms” An AI translation gave me: ❌ “voetbaluniformen” This is not how Dutch speakers actually say it. A professional translator chooses: ✅ “voetbaltenues” That’s the difference between literal translation and natural, market-ready language. AI can support the process. Humans deliver the result. Huisman Translations, Human Translations. #Translation #Localization #l10n #t9n #Translators #LanguageServices #EnglishDutch #PostEditing #HumanTranslations

Zero-shot translation meant translating between languages the AI never learned together. It worked beautifully until it didn't. The problem? AI models kept outputting text in the wrong language. Imagine asking for French translation and getting Spanish instead. Frustrating for users. Catastrophic for businesses. Researchers from China University of Petroleum and University of Sydney just cracked this code. Their breakthrough: a two-stage fine-tuning method that teaches AI to follow translation instructions correctly. Here's how it works: • Stage 1: Train the model on basic translation data • Stage 2: Introduce "instruction-conflicting samples" with wrong instructions • The AI learns to avoid generating outputs for these bad instructions The results speak volumes: 🎯 62.4% reduction in wrong-language outputs 📈 9.7 BLEU point improvement in translation quality ✅ No performance loss in other language tasks This isn't just a research win. It's a practical solution for real-world AI translation. Think about global customer service. Medical translation. Legal documents. Getting the language wrong isn't just inconvenient - it's dangerous. Zero-shot translation now works across 16 language pairs that weren't in the original training data. AI translation just got more reliable. And that changes everything for global communication. What language barriers do you face in your work? How could better AI translation help? #AITranslation #MachineLearning #GlobalCommunication 𝐒𝐨𝐮𝐫𝐜𝐞: https://lnkd.in/ej-7fZyJ

Zero-shot translation meant translating between languages the AI never learned together. It worked beautifully until it didn't. The problem? AI models kept outputting text in the wrong language. Imagine asking for French translation and getting Spanish instead. Frustrating for users. Catastrophic for businesses. Researchers from China University of Petroleum and University of Sydney just cracked this code. Their breakthrough: a two-stage fine-tuning method that teaches AI to follow translation instructions correctly. Here's how it works: • Stage 1: Train the model on basic translation data • Stage 2: Introduce "instruction-conflicting samples" with wrong instructions • The AI learns to avoid generating outputs for these bad instructions The results speak volumes: 🎯 62.4% reduction in wrong-language outputs 📈 9.7 BLEU point improvement in translation quality ✅ No performance loss in other language tasks This isn't just a research win. It's a practical solution for real-world AI translation. Think about global customer service. Medical translation. Legal documents. Getting the language wrong isn't just inconvenient - it's dangerous. Zero-shot translation now works across 16 language pairs that weren't in the original training data. AI translation just got more reliable. And that changes everything for global communication. What language barriers do you face in your work? How could better AI translation help? #AITranslation #MachineLearning #GlobalCommunication 𝐒𝐨𝐮𝐫𝐜𝐞: https://lnkd.in/ej-7fZyJ

Zero-shot translation meant translating between languages the AI never learned together. It worked beautifully until it didn't. The problem? AI models kept outputting text in the wrong language. Imagine asking for French translation and getting Spanish instead. Frustrating for users. Catastrophic for businesses. Researchers from China University of Petroleum and University of Sydney just cracked this code. Their breakthrough: a two-stage fine-tuning method that teaches AI to follow translation instructions correctly. Here's how it works: • Stage 1: Train the model on basic translation data • Stage 2: Introduce "instruction-conflicting samples" with wrong instructions • The AI learns to avoid generating outputs for these bad instructions The results speak volumes: 🎯 62.4% reduction in wrong-language outputs 📈 9.7 BLEU point improvement in translation quality ✅ No performance loss in other language tasks This isn't just a research win. It's a practical solution for real-world AI translation. Think about global customer service. Medical translation. Legal documents. Getting the language wrong isn't just inconvenient - it's dangerous. Zero-shot translation now works across 16 language pairs that weren't in the original training data. AI translation just got more reliable. And that changes everything for global communication. What language barriers do you face in your work? How could better AI translation help? #AITranslation #MachineLearning #GlobalCommunication 𝐒𝐨𝐮𝐫𝐜𝐞: https://lnkd.in/ej-7fZyJ

Zero-shot translation meant translating between languages the AI never learned together. It worked beautifully until it didn't. The problem? AI models kept outputting text in the wrong language. Imagine asking for French translation and getting Spanish instead. Frustrating for users. Catastrophic for businesses. Researchers from China University of Petroleum and University of Sydney just cracked this code. Their breakthrough: a two-stage fine-tuning method that teaches AI to follow translation instructions correctly. Here's how it works: • Stage 1: Train the model on basic translation data • Stage 2: Introduce "instruction-conflicting samples" with wrong instructions • The AI learns to avoid generating outputs for these bad instructions The results speak volumes: 🎯 62.4% reduction in wrong-language outputs 📈 9.7 BLEU point improvement in translation quality ✅ No performance loss in other language tasks This isn't just a research win. It's a practical solution for real-world AI translation. Think about global customer service. Medical translation. Legal documents. Getting the language wrong isn't just inconvenient - it's dangerous. Zero-shot translation now works across 16 language pairs that weren't in the original training data. AI translation just got more reliable. And that changes everything for global communication. What language barriers do you face in your work? How could better AI translation help? #AITranslation #MachineLearning #GlobalCommunication 𝐒𝐨𝐮𝐫𝐜𝐞: https://lnkd.in/ej-7fZyJ

Zero-shot translation meant translating between languages the AI never learned together. It worked beautifully until it didn't. The problem? AI models kept outputting text in the wrong language. Imagine asking for French translation and getting Spanish instead. Frustrating for users. Catastrophic for businesses. Researchers from China University of Petroleum and University of Sydney just cracked this code. Their breakthrough: a two-stage fine-tuning method that teaches AI to follow translation instructions correctly. Here's how it works: • Stage 1: Train the model on basic translation data • Stage 2: Introduce "instruction-conflicting samples" with wrong instructions • The AI learns to avoid generating outputs for these bad instructions The results speak volumes: 🎯 62.4% reduction in wrong-language outputs 📈 9.7 BLEU point improvement in translation quality ✅ No performance loss in other language tasks This isn't just a research win. It's a practical solution for real-world AI translation. Think about global customer service. Medical translation. Legal documents. Getting the language wrong isn't just inconvenient - it's dangerous. Zero-shot translation now works across 16 language pairs that weren't in the original training data. AI translation just got more reliable. And that changes everything for global communication. What language barriers do you face in your work? How could better AI translation help? #AITranslation #MachineLearning #GlobalCommunication 𝐒𝐨𝐮𝐫𝐜𝐞: https://lnkd.in/ej-7fZyJ