面谈
Olga Beregovaya, AI在Smartling的副总裁 – 访谈系列
在 Google 上的首选来源中添加 Unite.AI
Olga Beregovaya是Smartling的AI副总裁，她领导AI战略跨越翻译，数据，质量和安全。 她自1997年起从事自然语言处理，跨越基于规则的系统，统计模型，神经网络，以及大型语言模型。 在加入Smartling之前，她在Welocalize担任AI Innovation的副总裁和PROMT美洲公司的CEO.
Smartling提供企业翻译和本地化软件与服务，结合AI和机器翻译与人的专门知识。 其平台支持网站，应用程序，营销，文件，多媒体的多语言工作流程，可访问超过20个LLMs和机器翻译引擎。 在此次采访中，贝雷戈瓦亚讨论了多语种AI鸿沟，翻译质量，模型选择，以及人类监督在自动化本地化中的作用。
这段旅程如何塑造你对AI在语言中的作用的看法
我研究了语言的结构以及语言家庭的工作方式。 这告诉我，翻译是关于结构和意义，而不是字换字的替代。 我也这么想AI
我从1997年开始从事自然语言处理工作，当时系统是建立在手写规则之上的。 从那时起，我通过统计模型、神经网络和今天的大型语言模型开展工作。 每一个都解决了一些问题并创造了其他问题。 我认为没有一个模式是答案。
我加入Smartling是因为它是为了翻译各种企业内容而建造的，从网站和应用程序到法律文件，营销和视频。 我一开始是作为机器翻译的VP，但是这个角色迅速扩展，超越了翻译，变成了提示，数据，质量和安全。 因此这个头衔成为AI的VP.
在实践中，这种分裂是什么样子的，为什么随着基因AI成为主要接口而变得更加间接呢？
或支持你的语言，这个问题现在已基本解决，但分歧已经改变。 现在重要的是，该模式的培训数据如何很好地代表了该语言、语言和文化。 大多数模型训练数据都是英语，因此带有英语的假设。 译文在技术上可以准确，仍会将英语视角强加于另一种语言。
大赦国际正在成为人们与银行、医生、学校和其他基本服务机构互动的方式。 一个语言AI处理得很好让你有一个优势，一个语言处理不好让你处于劣势。
你最惊讶的是什么？
发言人人数是翻译质量的一个令人惊讶的差指标。 瑞典语有大约1 000万发言者，取得了一些最佳成果。 印地语人约6亿，与爱沙尼亚语和泰语一样，属于最弱者。
语言结构和数据质量更为重要。 瑞典语有相对简单的语法和丰富的清洁训练数据。 爱沙尼亚语不断修改词尾以传达意义，这为AI处理制造了更大的复杂性。 印地语有数亿的演讲者，但质量低得多，培训数据也不同，比从演讲者人数中预期的要少得多。
另一个惊喜是实验室的进展并不相同。 但这种模式并不总是真实的。
除了培训数据量之外，哪些因素对模型性能影响最大？
托肯化是最大的因素之一。 这对不使用拉丁字母的语言尤为重要。 糟糕的象征限制了接下来的一切。
语言结构也算数。 具有复杂词尾的语言可能更难让模型处理。 例如，德语将单词堆在一起，直到一个单词可以运行40个字符，这比西班牙语或法语等语言对一个模型提出了更大的要求。
还有数据质量。 不良数据产生不良结果。 数据的准确性、变异性和代表性对人工智能翻译的绩效至关重要。
最后，测量。 我们使用TER,BLEU,CHRF3和Metrix，并将结果与值得信赖的人类翻译进行比较。 一个模型在理解错误的同时可以听起来自然。
前沿实验室负有什么责任，有意义的进展是什么样子？
边疆实验室的责任最大，因为它们控制着最先进的语言模型。 他们的选择可以扩大鸿沟，或者帮助缩小鸿沟。 然而，翻译经常与推理，编码和安全争夺注意力，因此很少有实验室有专注团队关注。
三件事将产生有意义的影响：
- 建立更多语言和科目的高质量数据。
- 改进目前方法不足的语文的标志化。
- 衡量每个新模式与培训数据较少的语言，不只是最大的语言。
一个主要的机会是为网上代表很少的语文建立培训数据。 大部分工作发生在大实验室之外我们与非洲语言实验室等团体合作，为非洲语言建立数据。
Smartling如何确定其20+引擎中哪一种适合特定语言，内容类型，或使用大小写？
我们测试它们而不是假设一个模型是最好的。 每个模型都使用TER、BLEU、CHRF3和Metrix等可信赖的人类翻译进行评价。 一个客户可以给我们它的营销内容，大约一周内，我们可以显示每个模型如何按语言和主题进行。 拉丁美洲西班牙语营销的最佳模式可能不是另一种语言或敏感内容的最佳模式。
我们还考虑到风险。 低风险用户内容可以通过较轻的审查程序。 规范或品牌批判内容获得更多评论，通常为人检查。 因为模型可以随着性能变化而添加或删除
专业翻译系统在哪些方面仍然优于通用的LLMs，两者如何趋同？
在可预测性最为重要时，专门机器翻译仍然获胜。 它对敏感的内容和某些语言特别有用。 LLMs产生更多的自然写作，但是它们也可以有自信地产生幻觉。
在Smartling，我们结合了两者。 专业翻译模式利用客户的术语和域数据制作初稿。 LLM然后编辑它，恢复批准的术语，应用以前翻译的匹配，并改进写作。
这个模型正成为系统的一部分。 现在，提示，数据，定制和安全控制同样重要。
在使用基因AI处理多语种内容时，企业应当考虑的最大风险是什么？
幻觉，双手放下。 我最喜欢的例子是有人问人工智能蘑菇是否有毒。 并用个人墓碑证明这是错误的。 在翻译中，这个显示作为写作，听起来很完美，同时把事实弄错了，在培训数据较少的语言中，这种风险可能更高。
还有文化和监管背景。 翻译在技术上可以正确，对于市场来说仍然是错误的。 例如，德语和日语需要英语所没有的正规程度。 中国广告规则也限制声称，如"最佳".
一致性是另一个。 在规模上，通用AI工具可以产生不一致的术语和品牌语音。 一位执行官告诉我他的公司有30万雇员，由于通用AI,30万个不同的声音。
而利害可有形。 如果关于拉杆的安全指示被错误翻译，起重机可能会掉到施工现场。
AI agent在本地化中是什么样子，在人的监督下可以实现多少自主？
AI agent是本地化工作流程中的一个步骤。 一种模式翻译，另一种模式审查，检查幻觉，恢复经批准的术语，并为市场调整内容。 然后进行质量检查，决定能否公布，需要快速的人文审查，还是需要全面审查。
有多少可以自动化取决于风险。 我们的低风险工作大部分已经完全自动化，但敏感和规范的内容仍然需要人。
完全自动化并不总是目标，目标是减少人类的工作量，而不消除人类的判断力。
五年后，这一差距会缩小还是扩大？是什么使大赦国际在全球真正多语种？
不管是哪条路都行开始时最弱的语言正在迅速改进，更多的组织正在收集网上内容很少的语言的数据，包括口语数据。
但鸿沟正在变形。 AI能够很好地处理语言，同时仍然从其培训数据中导入英语假设。
真正的多语种AI需要不同的方法。 这可能意味着适合具体语言和主题的较小模式，而不是一个预期能做任何事情的一般模式。 这意味着不同语言和领域的数据更加多样化。 更好地标注目前方法不足的语文。 合成数据，加上一个模型从一种语言学到的东西转移到相关语言，所以改进并不限于数据最多的语言。
该技术存在，但问题是这些模型的建设公司是否将多语种的绩效列为优先事项。
感谢您的伟大采访，希望探索AI翻译，多语种质量评价，以及具有人类监督的本地化工作流程的读者应该访问Smartling.
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Interviews
Olga Beregovaya, VP of AI at Smartling – Interview Series
Add Unite.AI to your preferred sources on Google
Olga Beregovaya is Vice President of AI at Smartling, where she leads AI strategy across translation, data, quality, and safety. She has worked in natural language processing since 1997, spanning rule-based systems, statistical models, neural networks, and large language models. Before joining Smartling, she served as VP of AI Innovation at Welocalize and CEO of PROMT Americas.
Smartling provides enterprise translation and localization software and services, combining AI and machine translation with human expertise. Its platform supports multilingual workflows for websites, applications, marketing, documents, and multimedia, with access to more than 20 LLMs and machine translation engines. In this interview, Beregovaya discusses the multilingual AI divide, translation quality, model selection, and the role of human oversight in automated localization.
How has that journey shaped the way you think about AI’s role in language, and what drew you to lead AI strategy at Smartling?
I studied how languages are structured and how language families work. That taught me that translation is about structure and meaning, not word-for-word substitution. I think about AI the same way.
I started working in natural language processing in 1997, when systems were built on hand-written rules. Since then, I’ve worked through statistical models, neural networks and today’s large language models. Each solved some problems and created others. I don’t believe any one model is the answer.
I joined Smartling because it was built to translate every kind of enterprise content, from websites and apps to legal documents, marketing and video. I started as VP of Machine Translation, but the role quickly expanded beyond translation into prompting, data, quality and safety. So the title became VP of AI.
What does that divide look like in practice, and why could it become more consequential as generative AI becomes a primary interface?
The divide used to be simple: either an AI system supported your language or it didn’t. That question is mostly settled now, but the divide has changed shape. What matters now is how well the model’s training data represents that language, its speakers and its culture. Most model training data is in English, so it carries English-speaking assumptions with it. A translation can be technically accurate and still impose an English-speaking perspective on another language.
AI is becoming the way people interact with banks, doctors, schools and other essential services. A language AI handles well gives you an advantage, and a language it handles poorly puts you at a disadvantage.
What surprised you most, and what explains cases where Swedish outperforms Hindi despite far fewer speakers?
Speaker count is a surprisingly poor indicator of translation quality. Swedish, with about 10 million speakers, produced some of the best results. Hindi, with about 600 million speakers, was among the weakest, along with Estonian and Thai.
Language structure and data quality matter much more. Swedish has relatively simple grammar and plenty of clean training data. Estonian constantly changes word endings to convey meaning, which creates more complexity for AI to handle. Hindi has hundreds of millions of speakers, but far less high-quality, varied training data than you might expect from its speaker count.
The other surprise was that progress wasn’t the same from lab to lab. Languages that started out weaker usually improved the most from one model generation to the next, but that pattern doesn’t always hold true.
Beyond the volume of training data, what factors most affect model performance?
Tokenization is one of the biggest factors. It’s how a model breaks text into pieces before processing it, which is especially important for languages that don’t use the Latin alphabet. Poor tokenization limits everything that follows.
Language structure counts too. Languages with complex word endings can be harder for models to handle. Or, for example, German stacks words together until a single word can run 40 characters, which places greater demands on a model than languages like Spanish or French.
Then there’s data quality. Bad data produces bad results. The accuracy, variation and representation of data is incredibly critical for AI translation performance.
And finally, measurement. We use TER, BLEU, CHRF3 and MetricX, and compare results with trusted human translations. A model can sound natural while getting the meaning wrong.
What responsibility do frontier labs have, and what would meaningful progress look like?
Frontier labs have the most responsibility because they control the most advanced language models. Their choices can either widen the divide or help narrow it. However, translation often competes with reasoning, coding and safety for attention, so few labs have dedicated teams focused on it.
Three things would make a meaningful difference:
- Build high-quality data across more languages and subjects.
- Improve tokenization for languages where current approaches fall short.
- And measure every new model against languages with less training data, not just the biggest languages.
A major opportunity is building training data for languages that have very little representation online. Much of that work happens outside the big labs. We work with groups like the African Language Lab to build data for African languages.
How does Smartling determine which of its 20-plus engines fits a given language, content type, or use case?
We test them instead of assuming one model is best. Every model is evaluated against trusted human translations using TER, BLEU, CHRF3 and MetricX. A customer can give us its marketing content, and within about a week we can show how each model performs by language and subject. The best model for Latin American Spanish marketing may not be the best model for another language or for sensitive content.
We also factor in risk. Lower-risk user content can go through a lighter review process. Regulated or brand-critical content gets more reviews and usually a human check. Because models can be added or removed as performance changes, we aren’t locked into one provider.
Where do specialized translation systems still outperform general-purpose LLMs, and how will the two converge?
Specialized machine translation still wins when predictability matters most. It is especially useful for sensitive content and certain languages. LLMs produce more natural writing, but they can also hallucinate with confidence.
At Smartling, we combine the two. A specialized translation model produces the first draft using the customer’s terminology and domain data. An LLM then edits it, restoring approved terminology, applying matches from previous translations and improving the writing.
The model is becoming just one part of the system. Now, prompting, data, customization and safety controls matter just as much.
What are the biggest risks enterprises should consider when using generative AI for multilingual content?
Hallucinations, hands down. My favorite illustration is someone asking an AI whether a mushroom is poisonous. It answers no with complete confidence, and the person’s tombstone proves it wrong. In translation, this shows up as writing that sounds perfect while getting the facts wrong, and that risk can be higher in languages with less training data.
Then there’s cultural and regulatory context. A translation can be technically correct and still be wrong for the market. German and Japanese, for example, require levels of formality that English doesn’t. Chinese advertising rules also restrict claims such as “the best.”
Consistency is another one. At scale, generic AI tools can create inconsistent terminology and brand voice. One executive told me his company has 300,000 employees and, thanks to generic AI, 300,000 different voices.
And the stakes can be physical. If a safety instruction about pulling a lever is mistranslated, a crane can fall on a construction site.
What does an AI agent look like in localization, and how much can become autonomous with human oversight?
An AI agent is one step in a localization workflow. One model translates, while another reviews, checks for hallucinations, restores approved terminology and adapts the content for the market. A quality check then decides whether it can publish, needs a quick human review or needs a full review.
How much can be automated depends on the risk. Much of our lower-risk work is already fully automated, but sensitive and regulated content still requires humans.
Full automation isn’t always the goal, because models need human feedback to improve. The goal is to reduce the human workload without removing human judgment.
In five years, will the gap narrow or widen, and what would make AI genuinely multilingual at global scale?
It could go either way. Languages that started out weakest are improving fastest, and more organizations are collecting data for languages with little online content, including spoken-language data.
But the divide is changing shape. AI can handle a language well while still importing English-speaking assumptions from its training data.
Genuine multilingual AI requires a different approach. That could mean smaller models tailored to specific languages and subjects, rather than one general model expected to do everything. It means more varied data across languages and domains. Better tokenization for languages where current approaches fall short. And synthetic data, plus transferring what a model learns from one language to related ones, so improvements aren’t limited to the languages with the most data.
The technology exists, but the question is whether the companies building these models make multilingual performance a priority.
Thank you for the great interview, readers who wish to explore AI translation, multilingual quality evaluation, and localization workflows with human oversight should visit Smartling.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.