Alibaba International Launches New Large Language Model to Enhance E-Commerce Translation

(MENAFN- mslgroup) Alibaba International Digital Commerce Group ("Alibaba International") announces the launch of Marco-MT, the groundbreaking translation tool designed to break down language barriers with remarkable accuracy and intelligence.

Powered by Alibaba International's revolutionary large language model (LLM), Marco-MT intelligently interprets the nuances of context, breathing life into product descriptions, facilitating flawless multilingual interactions and precise document translation.

According to FLORES, a benchmarking evaluation framework for translation, Marco-MT has surpassed global leading AI translation products in terms of BLEU, an automatic evaluation metric.

FLORES Benchmark
BLEU Score Marco-MT Google Translate DeepL Translate GPT-4
Average 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

In evaluations for translation from other languages to English and vice versa, Marco-MT consistently delivers superior results.

Excelled in E-commerce Scenarios

Marco-MT seamlessly integrates into e-commerce systems, providing automated translation services to merchants and consumers. Its localization capabilities translate merchant-uploaded product information—including titles and descriptions—into the target market language, significantly enhancing communication efficiency and consumer outreach.

While other systems incorrectly translate the Chinese e-commerce term '秒杀' (flash sale) into the Spanish 'eliminación rápida de un adversario' (quickly kill your enemy), Marco-MT accurately translates it as 'Oferta flash'.

While peers tend to ignore one of the languages in a live-chat scenario with mixed languages, Marco-MT precisely translates them all into the target language.

Source Content Peer 1 Peer 2 Marco-MT
现在没有这个保护套هل سيتوفر قريبًا؟ There is no such protective case right now? This protective case is not available now هل سيتوفر قريبًا؟ Is this protective case available now? Will it be available soon?


How Is This Achieved?

Marco-MT benefits from Alibaba International's extensive experience and insights gained from multiple global e-commerce platforms, accumulating several billion e-commerce data entries. This vast data set, combined with sophisticated techniques for data filtering, advanced language identification, and quality evaluations, empowers Marco-MT to excel across a variety of translation metrics, including BLEU, COMET, and human evaluations.

The development of Marco-MT involved multi-language enhancement training through the Qwen series of large language models, ensuring a robust multilingual framework.

By employing innovations such as multilingual mixtures of experts (MOE) and parameter expansion methodologies, Marco-MT maintains top-tier performance in dominant languages, while simultaneously bolstering the capabilities of other languages.

Additionally, through optimization strategies like model quantization, acceleration, and multi-model reduction, Alibaba International significantly lowers the service costs of large models, making them more cost-effective than smaller models.

How Does It Benefit Users?

Currently, Marco-MT has successfully achieved large-scale commercial use, demonstrating exceptional performance in cross-border e-commerce, a result of Alibaba International's years of expertise. For individual users, Marco provides high-quality, contextually relevant translations and supports various styles to meet diverse needs. It is available online, and users can start trialing it through Aidge’s website.

Last year, Alibaba International established an AI team to explore capabilities across 40 scenarios, optimizing 100 million products for 500,000 small and medium-sized enterprises.

Demand for AI among merchants is rapidly increasing, with usage rates doubling approximately every two months, leading to over 100 million average daily AI calls. This growth underscores the e-commerce industry's reliance on AI tools, setting a new standard for business operations and customer engagement.

MENAFN07112024004993016390ID1108861193