news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Alibaba International used AI to rewrite the titles and detail pages of 100 million products.

Author: Wang Yi

Editor｜Li Zi

2023 is the year when cross-border e-commerce will make rapid progress overseas. The "Four Little Dragons" of Chinese e-commerce (SHEIN, AliExpress, Temu and TikTok Shop) have entered an era of rapid expansion and are competing fiercely with overseas rivals in the global market.

Under the "full trusteeship (merchants are only responsible for supply, and cross-border e-commerce platforms provide operations, logistics, after-sales and other services)" and "semi-trusteeship" (the platform is responsible for selling goods, but merchants bear the costs of international trunk logistics and delivery after arriving at the destination country) models launched by major e-commerce platforms, merchants' sales have achieved significant growth.

But apart from relying on the powerful supply chain and extreme cost-effectiveness of Made in China, how else can China achieve long-term sustainable growth and compete globally with overseas giants such as Amazon?

Click on "Technology Tree"Maybe it’s the only solution.

With the arrival of the generative AI era, the rapid iteration of AI tools, from ChatGPT to Sora, has also brought more possibilities to the cross-border e-commerce industry: not only are more than 80% of independent site sellers exploring and using AI technology to reduce costs, increase efficiency and improve user experience, but major cross-border e-commerce platforms have also been deploying AI since the second half of 2023.

For example, Amazon Web Services launched an artificial intelligence service called Amazon Bedrock last year, allowing customers to build GenAI applications through Anthropic, Stability AI and Amazon's existing models; Shein uses AI to predict fashion trends and generate e-commerce images; Shopee, Lazada, TikTokShop, Shopify and others are also successively deploying AI e-commerce.

A case that better proves the commercial value of AI in the e-commerce field is Alibaba International Digital Business Group (hereinafter referred to as "Ali International").

In Alibaba's first quarter 2024 financial report, Alibaba's international retail business revenue was 22.278 billion yuan, a year-on-year increase of 56%, making it the business with the largest growth rate this quarter.

The "dark horse" Alibaba International did not emerge recently.

In the first three quarters of 2022, Alibaba's international retail business growth was still in the single digits, and even experienced a slight decline and fluctuation due to the epidemic. Since the second quarter of 2023 (natural season of natural year), international retail has begun to soar - not only maintaining a high double-digit year-on-year growth, but also the growth rate has become steeper; in the four quarters of 2023, the year-on-year growth rates were 41%, 60%, 73%, and 56%, respectively, making it Alibaba's fastest growing business in the past three years.

Behind Alibaba International’s rapidly growing performance, there is a secret driving force: AI.

1. E-commerce going globalAI"shipbuilding"

In April 2023, Alibaba International established an internal "AI Business" team in an attempt to empower cross-border e-commerce with AI capabilities.

The reason for trying to use AI to solve cross-border e-commerce problems is related to the environment in which Alibaba International itself is located.

As the domestic involution becomes more serious, Chinese Internet companies are going overseas. As the first Chinese company to carry out cross-border e-commerce business, Alibaba has the first-mover advantage, but it also has to beware of powerful competitors all around, and cannot relax in the supply of product categories, logistics delivery timeliness, after-sales service guarantee and other links.

At the same time, the characteristics of cross-border e-commerce business also determine that this is a field that is very suitable for upgrading and transformation using AI:

• First, the overseas market has diverse languages ​​and cultures, complex regulatory compliance requirements, and different countries have different time zones, which poses great challenges to merchants in writing product names and detail pages, arranging customer service working hours, etc.

• Secondly, in terms of marketing and user acquisition, overseas markets are relatively more difficult and the cost of acquiring customers is relatively higher;

• Third, there is still a severe shortage of talent in cross-border business, and positions such as design, finance, legal affairs, and customer service are in urgent need of professional talents.

All of the above are the "ninety-nine eighty-one difficulties" faced by merchants engaged in cross-border e-commerce, especially small and medium-sized merchants, on the road to going overseas.

The AI ​​big models that have emerged in recent years provide an excellent solution to the above problems.

Taking marketing as an example, AI can automatically generate multiple content based on the characteristics of users and products. It can also quickly and batch generate a large number of advertising materials, and generate a large amount of SEO text descriptions, selling points, etc. to help increase traffic.

In the shopping guide experience, AI can not only generate real-life model upper body pictures to replace clothing flat pictures, but also rewrite the main picture's problems such as "many psoriasis, insufficient clarity, unclear display", and enhance the beauty and clarity of the main picture;

In addition, in response to the translation problem that is a headache for most merchants and users, the multilingual AI model can provide accurate translation and error correction services in small languages, so that language is no longer a barrier; and in the after-sales stage, AI can also identify the degree of damage and scratches of the goods, etc., to help users recover their losses.

Previously, there has been a rumor that "AI is useless", mainly criticizing the difficulty of commercializing large AI models. Sequoia Capital also stated at the AI ​​Ascent at the beginning of the year that in 2023, the US technology industry spent $50 billion on GPUs for training AI, but the annual revenue of generative AI was only $3 billion, with an input-output ratio of 17:1.

However, the main reason why large AI models are not profitable is not the technology, but the inability to find suitable application scenarios.

Baidu founder Robin Li once publicly stated:"The model itself does not directly generate value. The application developed based on the basic large model is the meaning of the model. For entrepreneurs, the large model is meaningless, and the large application has greater opportunities."It seems to be a consensus among Internet companies to reversely infer the required AI capabilities based on application scenarios and then consolidate the foundation from multiple aspects such as data, algorithms, and computing power.

When AI, which was struggling to find application scenarios, encountered cross-border e-commerce, which was in urgent need of improving quality and efficiency, supply and demand were highly matched, and the two quickly hit it off. Alibaba International AI Business was born.

2. Alibaba InternationalAIaccelerate

Currently, Alibaba's AI Business team has a team size of more than 100 people, of which about one-third are algorithm R&D engineers, and the rest are responsible for model reasoning applications, infrastructure, and the development and operation of specific products.

The algorithm team of AI Business focuses on three directions:Multi-language e-commerce large model training; dialogue models and downstream tasks; image generation and understanding capabilities,They are formulated to meet the needs of different scenarios such as product information localization, customer service robots and customer service translation, content seeding and advertising placement in cross-border e-commerce.

Relying on the strong technical capabilities of the AI ​​Business team, Alibaba International has been continuously experimenting and exploring the application side of cross-border e-commerce over the past year, using AI to optimize more than 40 scenarios such as advertising, AI image generation, shopping guide experience, and refund agents:

for exampleAdvertisingIn order to solve the pain points of long production cycle and production bottleneck of manual marketing materials, Alibaba International has developed an AI full-link advertising creative material generation solution with automatic product selection, selling point extraction and image generation through AI capabilities such as multimodal recognition, multilingual text generation, AI image processing and delivery effect reinforcement learning. It can not only intelligently generate marketing advertising materials for a product in batches, but also learn excellent delivery cases, generate marketing diagram templates, and generate scoring models based on training data to select the best marketing materials. After adding AI capabilities, the advertising production cost of merchants has dropped by 5%, and the ROI of advertising has increased by 5%.

AI automatically generates marketing materials to improve advertising effectiveness and reduce advertising costs. Image source: Alibaba International

for exampleAIImage GenerationIn order to address the pain points of high production costs, time-consuming design workflows, and lack of international design and minority language talent in e-commerce, Alibaba International provides tools such as white-background images, marketing images, detail page translation, model images, model try-ons, video translation, and digital humans. These tools cover the entire scenario from product listing to off-site marketing material generation, helping merchants solve the problem of creating e-commerce images and videos from scratch.

AI provides virtual models with body effects for clothing categories. Image source: Alibaba International

Take the AliExpress seller Zeuslap as an example. This is an e-commerce brand specializing in selling monitors. Since November 2023, they have started using Alibaba International's AI image generation function for store decoration banners, product scene pictures, product details pictures, and cover pictures of major social media.

Zeuslap product page source: AliExpress

AI image generation has helped Zeuslap's owners, Mary and her husband, save a lot of time. They used to spend a lot of time looking for materials and designing, but now they only need to select templates and upload products to complete the task, and the results are better. AI capabilities have also greatly accelerated Zeuslap's performance growth, allowing it to quickly grow from a mom-and-pop store to the number one brand in the platform industry.

Another exampleRefund ScenarioIn response to the high cost of shipping goods back from overseas and the great damage to goods caused by failed returns, Alibaba International uses multimodal large model technology to identify dispute reasons and verify vouchers. After understanding various data, it calculates the amount of partial refund and provides consumers with a "no return, partial refund" solution to maximize customer satisfaction and minimize merchant costs.

Specifically, the model can be trained to learn and understand data such as user dispute reasons, user messages, refund vouchers (degree of damage to goods), transaction snapshots, logistics routes, etc., and by learning historical manual case handling methods, it can make judgments on "whether to return the goods" and "if not, how much to refund".

AI Refund Agent Workflow Source: Alibaba International

FinallyChargeback scenarioFor example, this is a post-sales dispute that all cross-border merchants are afraid of encountering - the goods have been shipped, but the consumer refuses to pay. In the past, in order to conduct a chargeback defense, merchants needed to spend hours preparing information and evidence for each link. Many small and medium-sized enterprises can only "accept their bad luck" when facing chargeback disputes due to lack of relevant experience. Now, Alibaba International's AI chargeback defense agent can organize all the information in a few minutes, automatically generate a defense letter, and send it to overseas credit card institutions.

AI chargeback defense agent helps merchants reduce financial compliance losses. Source: Alibaba International

Test data shows that AI Chargeback Defense Agent can help Chinese cross-border merchants on Alibaba International’s various platforms recover RMB 20 million in losses in a year and protect the rights and interests of merchants.

It can be said that the exploration of cross-border e-commerce scenarios has enabled Alibaba International's AI technology to expand from a domestic market of 1.4 billion people to a global market of 8 billion people. It has also provided cross-border sellers with a more convenient, efficient, and low-cost operational experience, greatly enabling the business growth of thousands of cross-border merchants.

On July 16, at an AI-themed sharing session held by Alibaba International, Zhang Kaifu, vice president of Alibaba International Digital Business Group and head of AI business, revealed that Alibaba International has tested AI capabilities in more than 40 scenarios, empowered 500,000 small and medium-sized businesses, and optimized 100 million products; in the past six months,On average, merchants’ calls for AI double every two months."The smaller and medium-sized businesses are, the more they can benefit from the application of AI," said Zhang Kaifu.

3. The secret behind massive calls

How is such a huge number of calls achieved?

The secret lies in Alibaba's international AI capabilitiesControllableandAvailable。

For quite a long time,The poor controllability and usability of large models are the reasons why they cannot be quickly applied.Taking Stable Diffusion, a commonly used image processing function, as an example, details such as fingers and hair are often distorted, making the generated model or product images unusable directly.

In addition, due to the lack of specific industry and scenario data, and the insufficient generalization capabilities of some large models, the usability of large models in e-commerce scenarios is poor. For example, the translation is not accurate enough, or it is difficult to achieve "one size fits all" in advertising delivery.

At the communication meeting, Zhang Kaifu said that AI Business is not a team that trains basic models, but makes flexible choices from a business perspective. If there are good enough open source models on the market, and after optimization using SFT, RAG, etc., the results are found to meet business expectations, then the existing open source models can be used. If some scenario requirements cannot be met, such as encountering the above problems, then self-developed models will be used, or proprietary technologies will be used for further training and optimization.

How to conduct self-research and optimization? Alibaba International focuses on three things:

• First of all, it is innovativeMultilingual text generation technologyTaking translation as an example, Alibaba International previously used a small model for translation in order to achieve faster reasoning and understanding, but the service utilization rate was low, resulting in a waste of resources. Later, they used a large model to integrate all languages ​​into a unified large model, and found that fewer servers needed to be deployed than before, and the service utilization rate was also improved.

This unified large model is the latest one launched by Alibaba InternationalMulti-language enhanced large model MarcoPoloIt is based on massive high-quality multilingual data training (including 2.5T tokens of minority languages), provides models with different parameter specifications such as 8B/57B/72B, supports 128K long context, and achieves better results at a lower cost.

• The second is the launch ofMultimodal large model MarcoPolo-VLIt is based on the industry's original structured embedding alignment model (SEA) training and tuning, and can provide 7B/14B models, which can be applied to image recognition, optimization, completion, generation and other tasks in e-commerce scenarios.

The underlying technology of MarcoPolo-VL is developed by Alibaba International.Multimodal large model architecture Ovis.Different from other MLLMs (multimodal large models) that usually generate content directly in an unstructured manner after the visual encoder is projected through the MLP connector, Ovis draws on the text embedding strategy in the LLM (large language model) and introduces a learnable visual embedding table to convert continuous visual features into probabilistic visual tokens, and then obtain structured visual embeddings through multiple index weighting of the visual embedding table.

Ovis consists of three key components: visual tokenizer, visual embedding table and LLM. Image source: https://arxiv.org/pdf/2405.20797

This new model architecture avoids the common limitations of MLLM based on MLP connectors, such as information loss, information error, and overfitting risk, and can better recognize and generate multimodal content.Evaluation results show that, at the same parameter level, Ovis outperforms mainstream open source MLLM in multiple benchmarks, and Ovis-14B outperforms the closed-source model Qwen-VL-Plus overall; Ovis-7B also performs well in multiple multimodal tasks such as visual perception, reasoning and programming, mathematics and science, and life scenarios.The model has also been fully open sourced in communities such as Huggingface and Github。

Image source: https://arxiv.org/pdf/2405.20797

In addition to self-developed models, Alibaba International will continue to improve the generation of multimodal content through a series of technical means. Taking the current biggest pain point of Wenshengtu, "poor controllability", as an example, in order to solve this problem, the algorithm team of AI Business will take two steps. The first step is to continuously improve the model capabilities through training, RAG, etc., and the second step is to gradually improve the output effect through multiple rounds of interaction.

The reason for "multiple rounds of interaction" is that most of the interactions between users and AI products are conducted through natural language, which is not precise enough, and the algorithm team needs to gradually optimize it in the background through parallel splitting (Pipeline Parallelism). For example, if a user tells AI to "put the lamp on the left side of the picture", this is a vague concept, but by slicing the picture and turning it into a nine-square grid, putting the lamp in the lower left corner of the table, and then gradually adjusting it through multiple rounds of interactive dialogue, it is more likely to generate a picture that satisfies the user.

Based on these technical optimizations, the visual and multimodal models developed by Alibaba International have achieved more stable and controllable image quality and more efficient image output speed. A picture can be generated in 8 seconds, and it can also automatically beautify the background and perform virtual fitting with one click, which greatly reduces the merchant's image production costs and increases the click-through rate of product images.

Image source: Alibaba International

• The third is to build a generalModel ServingPaaSPlatformSpark. It is based on Alibaba Cloud and other underlying infrastructure, and includes modules such as App Builder, Model Space, AI application engine, and AI online service platform. These modules together form a complete end-to-end technology chain (from training, reasoning to application), allowing all scenarios of e-commerce to share AI infrastructure. As mentioned earlier, the model workbench includes not only self-developed models, but also open source models and models of partners. Alibaba International will train, tune, evaluate, and deploy these models in a unified manner through prompt engineering, instruction fine-tuning, and other means.

After more than a year of hard work and experimentation, Spark's capabilities have been applied in more than 40 application scenarios, empowering 500,000 small and medium-sized businesses, and more than 100 million items of product information have been optimized.

At the same time, thanks to this platform, the use of AI in Alibaba International has begun to increase significantly. Data from the past six months show that on average, the number of merchants calling and reasoning on AI doubles every two months, and the reasoning cost can be reduced by 50% every 4-5 months, bringing a 1%-30% increase in click-through rate, conversion rate and consumer satisfaction for small and medium-sized merchants.

Alibaba International AI achievements over the past year. Image source: Alibaba International

4. Cross-border e-commerce “All in AI”

As mentioned above, 2023 is the first year of "AI e-commerce". In addition to Alibaba, Shein, TikTok Shop, Amazon, Shopify and other cross-border e-commerce platforms are also increasing their investment and deployment of AI:

In addition to launching Amazon Bedrock, a fully managed generative AI service that allows users to build applications, Amazon has also launched Amazon Q, a generative AI assistant that can help customers with software development and data analysis. At the same time, Amazon's AI Listing feature enables sellers to write higher-quality copy, and its latest e-commerce AI search tool Rufus can also provide buyers with purchasing suggestions, product comparisons, product recommendations and other content, shifting the original AI empowerment of e-commerce from the "conversion rate end" to the "traffic end". In addition, since 2023, Amazon has also launched a series of AI products around the three functions of AI-generated review highlights (AI review integration), Fit Review Highlights feature (AI clothing fit function) and generative AI to make product listings even more informative for customers (AI link writing), helping merchants improve the quality and efficiency of their operations.

Rufus product image, source: Internet

Shein has integrated AI and big data into its corporate genes since the first day of its focus on the "small order, quick return" flexible supply chain; after the advent of the generative AI era, Shein has used AI functions in image generation, image optimization, virtual fitting, customer service robots and other fields.

Social media rising star TikTok is making efforts in the advertising field and launched a dubbing tool called "Symphony AI", which can seamlessly translate content into more than a dozen languages ​​and dialects, allowing creators and businesses to cross cultural and language boundaries and effectively communicate with global audiences and consumers; they are also testing a new feature called "Symphony Digital Avatars", which allows brands to use AI to generate digital avatars during the advertising process to increase the "human touch" of advertising and improve brand marketing effectiveness and purchase conversion rates.

Schematic diagram of AI-generated advertising, source: Ad Age

The independent e-commerce platform Shopify has achieved a comeback with its AI strategy: they not only used AI to reconstruct 90% of the shopping experience on the website, but also launched the AI ​​merchant assistant SideKick, which can answer core business questions, adjust the website sales strategy and execute it, and quickly renovate and upgrade the website according to the seller's promotional needs. After launching the AI ​​function, Shopify's total revenue in the third quarter of 2023 reached US$1.71 billion, and its net profit successfully turned losses into profits.

Among all cross-border e-commerce platforms that have deployed AI, Alibaba International has its unique advantages:

The first is the advantage of e-commerce experience.

Alibaba International is one of the earliest Chinese cross-border e-commerce companies to go global. Backed by Alibaba, a huge e-commerce ecosystem, Alibaba International has more than ten years of e-commerce experience and more than 300 million overseas consumers. This has enabled it to accumulate rich insights into cross-border e-commerce scenarios and user behavior, providing the most natural and useful materials for large model training.

The second is industry know-how.

E-commerce is largely an experience-driven industry. Alibaba started cross-border business in 1999, and now has a history of 25 years. Alibaba International also includes AliExpress, Tmall Taobao Overseas, Lazada, Trendyol, Daraz, Miravia and other platforms, with a very rich reserve of e-commerce talents and business experience. These have accumulated a lot of industry know-how for it, making it smoother and less resistant when building AI-oriented business scenarios;

Finally, it is perfectAIinfrastructure.

Cross-border e-commerce has many business processes and complex scenarios, which require a large number of agents to assist in achieving automated workflows and thus reducing costs and increasing efficiency. Alibaba International has always adhered to the development strategy of "application first" and has long integrated AI and digital capabilities into existing business scenarios and product systems in a loosely coupled manner, retaining a large number of API interfaces, which enables it to quickly keep up with the pace after the arrival of the big model era and integrate the capabilities of the big model into various businesses.

In Alibaba International's view, compared with domestic e-commerce,AIThe large model provides greater support for cross-border e-commerce and the effect is more obvious.

Luo Weihua, head of the algorithm for Alibaba's international AI business, said that taking the "one thousand faces for one thousand people" product recommendation as an example, traditional algorithms need to have more knowledge and understanding of user information before they can complete this task. The current situation is that the user and order data of the domestic Taobao system is relatively rich, while the user data of overseas users is relatively sparse. "To achieve better results with sparse data, this can sometimes be a challenge."

With the support of large AI models, the problem of inaccurate recommendations caused by insufficient user data can be greatly alleviated, greatly improving recommendation effects and user satisfaction.

Zhang Kaifu also said that the core task of Internet platforms has always been "search promotion (search, recommendation, advertising)". After the big model came out, many people were concerned about "how the big model can transform search promotion", but the progress in China is very slow. The reason is that the search promotion in China has been done very well - relying on massive user data and high-quality algorithms, Douyin's content recommendation has become the world's leading, and Taobao's e-commerce search is also the best in the world. When the original technology is already advanced enough, it is difficult for the big model to completely subvert the original search promotion model from the underlying architecture, and it can only be icing on the cake;

On the other hand, in cross-border e-commerce, since user data is relatively sparse and search promotion is not yet complete and mature, after the limited and sparse data is connected to the big model, the effect of search promotion is often more prominent, the user experience is better, and the conversion rate is more significantly improved.

In the eyes of Alibaba International, the combination of AI and cross-border e-commerce is just the first step. After using AI to maximize the "cost reduction and efficiency increase" of e-commerce, Alibaba International may explore more scenarios combining cross-border e-commerce and AI, such as cutting products forward, exploring more diverse AI e-commerce product forms under the general trend of shelf e-commerce evolving into live broadcast e-commerce and content e-commerce; cutting supply chains backward, subverting the traditional "domestic trade to foreign trade" model, and using China's digital infrastructure to upgrade the globalization of "e-commerce infrastructure" such as logistics and supply chains. "We think that this is just the starting stage. In fact, in the future business growth space, AI can produce more innovative things," said Xu Zhao, head of Alibaba International's technology platform.

Next, Alibaba International also plans to open its AI capabilities to merchants outside the Alibaba ecosystem and use AI to empower a broader cross-border e-commerce market.

Whether it is reducing costs and increasing efficiency, or boosting performance, we have seen all the new possibilities that AI brings to cross-border e-commerce. We are also waiting to see how AI and big models will be more closely integrated with cross-border e-commerce in the future, and how Alibaba International will upgrade its AI products and capabilities.

(Cover image source: Alibaba)