内容来源：https://aibusiness.com/responsible-ai/combatting-cultural-bias-in-the-translation-of-ai-models

内容总结：

谷歌云发布研究报告：生成式AI应用需跨越“文化鸿沟”，语言细微差别成全球部署关键挑战

随着全球企业加速布局生成式AI，如何选择首个落地场景成为关注焦点。谷歌云近期发布的研究指出，初期应聚焦于提升人类信息交互体验的领域，其中，克服AI模型的语言与文化偏见是实现全球规模应用的核心挑战。

尽管科技公司持续推出翻译模型，但其在捕捉文化细微差别方面仍存明显不足。AI偏见不仅体现在性别与种族层面，更日益凸显为对特定语言的系统性偏好。为应对此问题，谷歌、OpenAI等厂商已开发多语言翻译模型，例如谷歌于1月15日发布的TranslateGemma，支持55种语言及500种语言对译。然而，现有模型在理解口语中的文化语境时仍力有未逮。

企业AI平台供应商Articul8通过其LLM-IQ智能体评估系统发现，多数模型在“文化适应性”维度表现欠佳。该系统从流畅度与自然性、连贯性、文化规范、一致性及清晰度五个维度对模型进行分层评估。Articul8创始人兼CEO阿伦·苏布拉马尼安在接受专访时指出，这一评估框架的诞生源于其客户在日韩市场遭遇的实际困境。

“在日本和韩国，客户反馈称：‘AI的回答准确，但显得无礼。’”苏布拉马尼安举例说明，日语中存在复杂的敬语体系，对话的语境、对象及预期结果都会影响表达方式的选择，而当前主流模型大多基于英语或拉丁语系数据训练，即使包含日语数据，也未能有效学习这些社交语言规则。

这种文化不适应性可能带来实际风险。在供应链管理或工业场景中，若AI生成的建议无法清晰传达指令的紧迫性或礼貌层级，可能导致人类操作者误判，进而引发重大损失。

针对“为何需要跨国企业而非本地厂商解决语言问题”的质疑，苏布拉马尼安强调“全球视野与本地化深耕相结合”的重要性。他以Articul8与Meta、AWS的合作为例，指出全球数据与合作伙伴网络能为本地化模型提供更丰富的训练基础，但最终仍需构建专属于目标语言文化的定制化模型。

究其根源，当前AI模型难以把握语言细微差别的主因在于训练数据的严重偏差。“英语与非英语内容的数据分布极不均衡，比例可能接近99:1，”苏布拉马尼分析称，“即使是数字化的非英语内容，也大多源自西方或特定封闭来源，导致模型难以学习多元文化中的交互规范。”

为突破这一瓶颈，Articul8采用“模型网格”架构，在运行时动态协调多个专用模型，而非依赖单一通用模型。该方案既利用通用模型获取世界知识，又针对特定语言（如日语）训练独立模型，通过系统化协作提升整体表现，避免了“为每项任务重建庞大模型”的资源浪费。

研究表明，生成式AI的全球化部署不仅需要技术迭代，更需深度理解语言背后的文化逻辑。企业若要在国际市场实现AI价值，必须将文化适应性纳入模型评估与优化的核心维度。

中文翻译：

谷歌云赞助
选择首个生成式AI应用场景
开启生成式AI之旅，首先应关注那些能够优化人类信息交互体验的领域。

尽管模型提供商正在推出翻译模型，但要确保它们精准捕捉文化细微差异，仍有大量工作有待完成。
AI偏见通常指大型语言模型有时对不同性别和种族表现出的系统性偏见，但如今越来越明显的是，模型也可能因偏好某种语言而产生偏见。
近年来，业界已努力遏制这种语言偏好，谷歌和OpenAI等AI模型开发者已创建翻译模型。最近，谷歌于1月15日发布了TranslateGemma，该模型使用55种语言和500种语言对（即易于互译的语言组合）进行训练。
然而，翻译模型未能捕捉口语中的某些细微差别。企业AI平台供应商Articul8表示，其LLM-IQ智能体在这方面能提供更深入的洞察。这个多层评估智能体系统从五个定性维度对模型进行评分：流畅性与自然度、连贯性、文化规范、一致性和清晰度。
通过该框架，Articul8发现许多模型在文化适宜性方面表现不佳，这表明要使AI技术在全球范围内成熟应用，仍需更多努力。
在本篇问答中，Articul8首席执行官兼创始人阿伦·苏布拉马尼扬探讨了该框架的开发背景，以及为何拥有文化适宜的模型至关重要。

是什么促使Articul8开发LLM-IQ智能体？为何专注于AI模型翻译的细微差别？
阿伦·苏布拉马尼扬：我们在日本和韩国拥有客户。当我们开始在这些地区部署时，我们需要真正理解多种语言的模型。
早期部署部分系统时，出现了一种情况：客户既满意又不满意。
在日本和韩国，他们告诉我们："你们的答案准确，但不够礼貌。"
我们回应说'好的'，但并不理解其中的差异。
事实证明，日语具有多层次的复杂性。许多语言都有类似特点。例如在英语中，"you"就是"你"，既无不敬也无特别尊敬之意。而在许多语言中，对平辈、长辈、上级或需表示尊敬的对象，使用的"你"是不同的词汇。这些细微差别有时能被捕捉，但多数时候则不能。
但日语还有更深一层：话语的语境，包括对话对象、说话者身份以及对话期望达成的结果。表达可以是直接、间接、礼貌、过度礼貌或稍显严厉的。根据语境，如果使用了错误的语调，也会被视为不当。
这确实引起了我们的兴趣，因为这更多是语言层面的问题。尽管不属于技术领域，但这是日语特有的领域语言。
经过深入研究，我们发现这具有系统性。所有模型主要基于英语或拉丁语系语言构建，甚至中国的模型也完全忽略了这种细微差别。它们在数字内容方面对日语的覆盖可能更广，但并未经过训练来捕捉这些细微之处。

LLM礼貌与否在哪些情境下至关重要？
苏布拉马尼扬：例如在供应链中，你无法分辨对方是在提出建议，还是在下达可能产生深远影响的指令。
这也可能带来严重的成本代价。
假设一个汽车系统生成了一条建议，参与决策的人正在阅读该建议。此人无法百分百确定是否需要执行该建议。这在工业环境中影响重大。

随着主权AI兴起，更多区域AI供应商用自有技术解决本地问题，为何应由日本以外的供应商来处理语言问题？
苏布拉马尼扬：我更倾向于将其视为具备全球视野者与仅具本地视野者的区别。你需要本地化赋能，同时保持全球视野。
关键在于将全球学习成果即时应用于日本，并进行独特的本土化适配。这很不同，因为虽然你能即时了解本地化需求，但想象一下在全球运营时，你需要处理完成工作所需的所有数据。
例如，我们的能源模型基于全球数据集，本地合作伙伴关系依托全球合作的制造模型。我们与Meta的研究合作、与AWS的扩展合作，都源于我们是一家全球运营商。但我们也深刻认识到，尽管业务全球布局，仍需对所做工作进行定制化。

为何您认为LLM似乎无法捕捉日语等语言的细微差别？
苏布拉马尼扬：最大的问题在于所有数据集都存在严重偏见。我所说的"偏见"是指英语与非英语内容的不对称分布。即使在拉丁语系或拉丁语源语言中，分布也是不对称的——接近99%比1%，并非微小差异。
甚至数字化的非英语内容也主要来自西方或我们无法获取的来源（如中国）。
所有关于礼貌与否、近乎自然人际互动的标准都源自西方。

在开发此框架过程中，是否有某个开源模型表现优于专有模型？
苏布拉马尼扬：我们对所有开源和闭源模型进行了基准测试。但随后我们必须从头构建这些模型，因为需要平衡数据集。如果不平衡数据集，将始终存在同样的偏见。
我们有一个称为"模型网格"的概念，使我们能够在运行时协调并决定调用哪个模型处理何种任务。我们不一定需要一个大型通用模型来为每项任务进行微调。我们可以拥有独立的任务专用模型，然后让它们作为系统协同工作。这样，系统就是一个可协同运行的实时推理引擎。
是的，我们确实使用通用模型来获取世界知识。但在涉及日本和日语时，我们使用自己的模型。
人们可能担心的另一个问题是："天哪，难道我需要为每项任务都构建庞大模型吗？"
答案是否定的，因为我们最终会形成一系列共同成长的模型。如果一个模型在某项任务上表现出色，其经验将以某种方式影响并全面提升整体表现。

编者注：为求清晰简洁，本次访谈内容经过编辑。

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
Although model providers are introducing translation models, much still needs to be done to ensure they accurately capture cultural nuances.
While AI bias has most often been the systemic bias large language models sometimes display against different genders and races, it is also becoming clear that models can be biased by a preference for one language over another.
In recent years, efforts have been made to curb this preference, with AI model developers such as Google and OpenAI creating translation models. Most recently, Google released TranslateGemma on Jan. 15, which was trained in 55 languages and 500 language pairs -- languages that can easily be translated from one to the other.
However, translation models fail to capture some of the nuances of spoken language. Enterprise AI platform vendor Articul8 says its LLM-IQ agent provides more insight into this. The multi-tiered evaluation agentic system scores models on five qualitative dimensions: fluency and naturalness, coherence, cultural norms, consistency and clarity.
With the framework, Articul8 found that many models failed on cultural appropriateness, suggesting that more work is needed for AI technology to be ready on a global scale.
In this Q&A, Articul8 CEO and founder Arun Subramaniyan discusses what led to the development of the framework and why having a culturally appropriate model is essential.
What led Articul8 to develop the LLM-IQ agent, and why did it focus on the nuances of translation in AI models?
Arun Subramaniyan: We have customers in Japan and in Korea. As we started deploying into those regions, we needed models that actually understood multiple languages.
One thing that happened was that when we deployed some of our systems early on, the customer was both happy and unhappy.
In Japan and Korea, they told us, "Your answer is accurate, but it's rude."
We said, 'Okay,' but we didn't know the difference.
It so turns out that Japanese has multiple layers of complexity. A lot of languages have that. For example, in English, you is only you. It's neither respectful nor disrespectful. Whereas in many languages there is a 'you' for people you're on the same level with, if you're addressing elders, seniors, or someone with respect, it's a different word. And those nuances are sometimes picked up, but most of the time they are not.
But in Japanese, there is one more level where the context of what you are saying, like who you are saying it to, who is saying it, and to get what outcome from that conversation. You can be direct, indirect, polite, overly polite, or slightly harsh. Depending on the context, if you use the wrong, say, intonation, that's also considered wrong.
That's really what intrigued us, because this is more at the linguistic level. Even though it is not a technical domain, it is a domain-specific language for Japanese.
After we did more research, we found it to be very systematic. All of the models were built predominantly with English or Latin-like languages, and even those from China missed this nuance completely. Their representation of Japanese in terms of digital content might be higher. However, they were not trained to catch these nuances.
In what situations would it matter whether an LLM is polite or impolite?
Subramaniyan: For example, in a supply chain, you don't know whether somebody was giving a recommendation or whether somebody gave a directive that will have profound implications.
Also, it might have serious costs.
If you have an automotive system, it is generating a recommendation. The human in the loop is reading the recommendation. The human doesn't know whether the recommendation needs to be acted on with 100% certainty. That has profound implications in an industrial setting.
With the rise of sovereign AI, with more regional AI vendors addressing local issues with their own technology, why should a vendor outside a country like Japan be the one to deal with the language problem?
Subramaniyan: I see this more as someone with global insight versus somebody with only local insight. You need to be locally enabled, but globally optimistic.
It's about global learning applied immediately in Japan, with localization that is uniquely Japanese. It's very different because, yes, you know more about the localization instantly, but imagine having to operate globally with all the data you need to do what you need to do.
For example, our energy models are based on global datasets. Our local partnerships are based on manufacturing models from global partnerships. Our research partnership with Meta, our scaling partnership with AWS, all of those come because we are a global operator. But we also operate with the deep understanding that even though we are global, we have to customize what we do.
Why do you think the LLMs appear unable to catch the nuance of a language like Japanese?
Subramaniyan: The biggest fault is that all of the data sets are extremely biased. What I mean by 'biased' is an asymmetric distribution of English versus non-English. Even in Latin languages or Latin-based languages, the distribution is asymmetric: I'm talking 99% to 1%. It's not like a slight difference.
Even digitized non-English content comes primarily from the West or from sources we don't have access to, such as China.
All that politeness, what is considered polite and impolite, what is considered near-natural human interaction came from the West.
In developing this framework, was there a particular open source model that worked better than proprietary models?
Subramaniyan: We benchmarked against all open source models and all closed source models. But then we had to build these models from the ground up because we had to balance the data set. If you don't balance the data set, you're going to constantly keep having the same bias.
We have a concept called Model Mesh, which enables us to orchestrate and decide at runtime which models to call for what. We don't necessarily need a large, general-purpose model that has to be fine-tuned for every task. We can have task-specific models that are independent and then make them work together as a system. Then the system is a runtime reasoning engine that we can run together.
Yes, we do use general-purpose models to acquire information about the world. But then, when it comes to Japan and the Japanese language, we have our own model.
The other question on people's minds would be, 'Oh my god, like do I need to build massive models for every single task?'
The answer is no, because we end up with a family of models that grow together. If a model does one task really, really well, that somehow influences and improves across the board.
Editor's note: This interview has been edited for clarity and conciseness