Introducing Marco-MT: Bringing Translation to the Next Level with LLM
Today, we proudly unveil Marco MT, Alibaba International's groundbreaking translation innovation fueled by our state-of-the-art Large Language Model (LLM).
Designed to transcend linguistic barriers with remarkable ease, Marco MT doesn't merely translate—it intelligently interprets the subtleties of context, breathing life into product descriptions, facilitating flawless multilingual interactions, and precision-document translation. Whether navigating the complexities of e-commerce terminology or enriching daily discourse, Marco MT sets a new benchmark, ensuring your global communications resonate with accuracy and authenticity.
According to FLORES, a benchmarking evaluation framework for translation, Marco MT has surpassed global leading AI translation peers in terms of BLEU, an automatic evaluation metric.
FLORES Score
Marco MT
Global Peer 1
Global Peer 2
Global Peer 3
General
47.70
43.53
42.90
39.52
Others to English
51.60
43.56
43.43
39.83
English to Others
43.79
43.50
42.36
39.21
How did we do it?
Benefitting from Alibaba International's years of insights accumulated across its multiple global e-commerce platforms, we have amassed a corpus of e-commerce data entries on a scale of several billion. In evaluations of e-commerce content translation, Marco MT outperforms leading AI translation products across metrics such as BLEU, COMET, and human evaluations.
Marco MT underwent multilingual enhancement training through the Qwen series of large language models, laying the groundwork for a robust, large-scale multilingual model. It acquires high-quality, extensive multilingual datasets through sophisticated multi-language data filtering techniques, which encompass advanced language identification and comprehensive data quality evaluations across multiple dimensions.
Furthermore, by integrating the Multilingual Mixture of Experts (MOE) and parameter expansion methodologies, Marco MT ensures that the performance of dominant languages like Chinese and English remains uncompromised while simultaneously boosting the capabilities of other languages.
User Scenarios
Marco MT primarily serves two user scenarios.
First, cross-border e-commerce. Marco MT can be effortlessly integrated into systems managing e-commerce products, customer service, and search guidance, offering automated translation services to merchants and buyers alike. This significantly bolsters consumer engagement and streamlines communication efficiency, thereby transforming the intricate landscape of cross-border e-commerce.
Marco MT facilitates content localization by precisely translating merchant-uploaded product information—including titles, descriptions, attributes, and image captions—into the target market's language. It also demonstrates commendable performance in translating search keywords and dialogue, capable of handling high concurrent invocations, delivering smooth and highly accurate translations that embody an understanding of e-commerce-specific knowledge and stylistic preferences.
Second, any non-e-commerce users with translation needs can access our text and image translation services as well. For individual users’ document translation needs, Marco MT provides high-quality translation with excellent readability, adaptable to various styles, catering to the diverse document translation needs of users across the world.
Currently, we primarily support translation among 15 languages: Arabic, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Turkish, and Ukrainian.
Going forward, we will continually invest in increasing the number of languages, enhancing translation quality across various scenarios, and consistently reducing the cost of services for users.
Last updated
Was this helpful?