Mozilla为Firefox带来基于本地计算机运算过程的官方免费翻译工具

Mozilla在Firefox浏览器中增加了一个官方翻译工具，需要注意的是，它不依赖云处理来完成工作，而是在用户自己的电脑上执行基于机器学习的过程。这对一个与Google和微软等巨头紧密联系的流行服务相比，在安全性和保护隐私的特征上看是一个巨大的进步。

这个翻译工具被称为"Firefox翻译"，可以在这里以扩展形式添加到Firefox浏览器：

https://addons.mozilla.org/en-US/firefox/addon/firefox-translations/

第一次翻译一种语言时，它需要下载一些支持性资源，如果需要，它可能会下载改进的语言模型，但实际的翻译工作最终是由电脑完成的，而不是在几百英里外的数据中心。一方面，目前它支持的语种也并不算多。

这一点很重要，不是因为很多人需要在离线时在浏览器中进行翻译--就像潜艇的屏蔽门一样，这不是一个真正有意义的用例--而是因为目标是减少对别有用心的云供应商的最终依赖，以完成不再需要他们资源的任务。

这个翻译工具实际上是欧盟资助的Bergamot项目的结果，Mozilla与几所大学合作开发了一套机器学习工具，使离线翻译成为可能。通常，这种工作是由数据中心的GPU集群完成的，在那里，大型语言模型（大小为千兆字节，有数十亿个参数）将被部署，以翻译用户的查询。

但是，虽然Google和微软（更不用说其他新晋竞争者）基于云的工具的表现开始愈发准确，而且（由于拥有近乎无限的计算能力）很迅速，但将要翻译的数据发送到第三方进行分析并送回，存在着基本的隐私和安全风险。对一些人来说，这种风险是可以接受的，而另一些人则希望在没有必要的情况下不要涉及这些互联网巨头。例如，如果Google翻译了小吃店的菜单，是否会开始成为香肠促销的目标？更重要的是，如果有人用已知的设备ID和位置来翻译移民或医疗文件，执法部门会来敲门吗？对于任何担心使用云供应商翻译的隐私问题的人来说，无论在什么情况下，做所有的离线工作都是有意义的。

This translation tool is called & quot;Firefox Translation & quot;, can be added here as an extension to Firefox browsers:

Https://addons.mozilla.org/en-US/firefox/addon/firefox-translations/

When translating a language for the first time, it needs to download some supporting resources, and if necessary, it may download an improved language model, but the actual translation is ultimately done by a computer, not in a data center hundreds of miles away. On the one hand, it does not support many languages at present.

This is important, not because many people need to translate in browsers when they are offline-- like the screen door of a submarine, this is not a really meaningful use case-- but because the goal is to reduce the ultimate reliance on cloud providers with ulterior motives to accomplish tasks that no longer need their resources.

This translation tool is actually the result of the EU-funded Bergamot project, and Mozilla has developed a set of machine learning tools in cooperation with several universities to make offline translation possible. Typically, this work is done by the GPU cluster in the data center, where large language models (gigabytes in size and billions of parameters) are deployed to translate users' queries.

But while Google and Microsoft (not to mention other new competitors) cloud-based tools are beginning to perform more accurately and quickly (because of their near-unlimited computing power), there are basic privacy and security risks when the data to be translated is sent to third parties for analysis and return. For some, the risk is acceptable, while others want to leave these Internet giants out of need. For example, if Google translates the menu of a snack bar, will it start to become a target for sausage promotion? More importantly, will law enforcement knock on the door if someone uses a known device ID and location to translate immigration or medical documents? For anyone worried about the privacy of using cloud provider translations, it makes sense to do all the offline work in any case.