翻译技术 | 阿里巴巴自主研发大语言模型 Marco MT 震撼登场！

文摘 2024-11-11 07:01 陕西

阿里巴巴推出一款“全新升级版”翻译工具

Alibaba Launches ‘Completely Revamped’ Translation Infrastructure

Hangzhou (China)-based e-commerce and tech conglomerate Alibaba has announced that it has released a proprietary large language model (LLM) that is “better than products offered by Google, DeepL, and ChatGPT.” Shots fired.

总部位于中国杭州的电商和科技巨头阿里巴巴宣布推出一款自主研发的大语言模型（LLM），声称“其性能超越谷歌、DeepL和ChatGPT等产品”。枪声已响，竞争趋于白热化。

The proprietary LLM — known as “Marco MT” — is to be deployed across Alibaba’s existing translation offering launched last year to small and medium-sized buyers and merchants looking to translate product listings, product descriptions, and other business communications.

这款名为“Marco MT”的自主研发大语言模型（LLM）将应用于阿里巴巴去年推出的翻译服务，面向中小型买家和卖家，用于翻译产品目录、产品详情介绍以及业务沟通。

Kaifu Zhang, International Vice President at Alibaba, said “we’ve had the [AI] system for about a year, and have half a million adoptions and more than 100M API calls per day. Basically, we’re seeing real value from our investment in AI.”

阿里巴巴国际部门副总裁张凯夫表示：“我们这个[人工智能]系统已运行一年左右，已有50万商户采用，每天API调用超过1亿次。我们在人工智能领域的投资已基本初见成效。”

“The announcement [of Marco MT] is a component of our AI stack. […] We just updated our translation infrastructure with a large language model. Previously it was based on the previous generation of natural language processing (NLP) and machine learning algorithms, but now we have completely revamped it,” he said.

“[Marco MT]是建立我们自己的AI堆栈的一部分……我们刚刚使用大语言模型更新了自身的翻译基础设施。之前这一基础设施立足于上一代自然语言处理（NLP）和机器学习算法，现在我们已经对其进行了全面升级。”他说道。

The update follows research led by Alibaba earlier this year, which measured neural machine translation output against large language model output for e-commerce content.

在此次升级之前，阿里巴巴于今年上半年进行了一项研究，对比了神经机器翻译与大语言模型在电子商务内容方面的翻译效果。

The company enables merchants to sell products worldwide by translating live chat conversations and product information — including text within images — into 29 languages, for USD 12 per 1 million characters. The solution includes optimized data for e-commerce, as well as glossary management and personalized customizations to the model’s output.

该公司可将实时聊天对话和产品信息（包括图片中的文字）翻译成29种语言，每100万字符收费12美元，助力卖家开展全球销售。这次升级包括优化电子商务数据、术语管理以及个性化定制翻译。

Eyeing International Expansion

着眼于国际扩张

The company is looking for international growthin several key markets, including the US, Europe, Brazil, and the Middle East following investments in Trendyol.

继投资Trendyol之后，阿里巴巴正寻求在多个关键国际市场进行扩张，包括美国、欧洲、巴西和中东。

Zhang explained, “A lot of our merchants come from China, Southeast Asia, or Turkey, and when they’re trying to sell to Europe or the Middle East, they face challenges in terms of language barriers, understanding the local market, having the right access to talent, and [producing] good enough marketing messages. All of this translates into conversion rates for those merchants and into the bottom line.”

张凯夫解释说：“我们许多商家来自中国、东南亚或土耳其，他们一直努力将产品销往欧洲或中东，但在此过程中，他们面临语言障碍、不了解当地市场、人才匮乏以及无法[产出]抓人眼球的营销信息等挑战。这些因素都会影响商家的转化率和利润。”

“By applying our language model, [merchants] get better translation quality, especially for some of the less represented languages. And that’s where the traditional and conventional NLP tools fail to deliver,” he added.

“使用我们这个大语言模型，[商家]可以获得更好的翻译质量，特别是针对一些小语种。而这正是传统的自然语言处理工具无法实现的。”他补充道。

Zhang explained to analysts how the success of merchants worldwide ultimately benefits Alibaba: “Our business model is driven by commission and advertising. […] When our merchants make more money, we get better revenue,” he concluded.

最后，张凯夫向分析员解释了全球商家的成功最终如何惠及阿里巴巴：“我们的商业模式依靠佣金和广告收入……商家赚的钱越多，我们也会获利更多。”

特别说明：本文内容选自Slator官网，仅供学习交流使用，如有侵权请后台联系小编删除。