How Large‑Model AI Transforms Multilingual E‑Commerce Search

Multilingual Search Challenges

Alibaba’s international e‑commerce platforms serve users in over 100 countries and more than 30 languages. Engineers and product teams are primarily Chinese speakers, leading to limited expertise in many low‑resource languages and difficulty maintaining high‑quality multilingual query understanding and precise matching.

ME‑LLM Construction

Powerful multilingual large‑language models (LLMs) such as Qwen and Llama are fine‑tuned on e‑commerce data to create a multilingual e‑commerce base model called ME‑LLM (Multilingual E‑Commerce LLM). The model is applied to the entire search pipeline: query understanding, product understanding, and relevance ranking.

Generative Semantic Relevance

Traditional discriminative relevance models are reformulated as generative models. A prompt concatenates a multilingual query and a product title; the LLM outputs a relevance label. Supervised fine‑tuning (SFT) on manually annotated relevance data yields a noticeable lift over multilingual BERT‑style models such as RoBerta. Further fine‑tuning of the domain‑adapted ME‑LLM adds additional gains. Because the LLM is large, knowledge is distilled into a smaller BERT‑style student model using massive log data, preserving most of the performance while reducing inference cost.

Relevance Reasoning with Chain‑of‑Thought

A relevance‑reasoning model mimics human logical inference: it first reasons about category, brand, and attribute intents before judging relevance. Prompted chain‑of‑thought (CoT) inference outperforms non‑reasoning baselines on a difficult test set, confirming the value of explicit reasoning.

Query Rewriting via RLSF

The semantic gap between user queries and product titles is addressed with a three‑phase architecture:

LLM‑QR‑SFT : SFT of ME‑LLM on manually labeled Query-Query pairs produces a query‑rewriting model.

QR‑RM (Reward Model) : Generates scores for rewrites based on relevance, recall gain (metric Sdelt_recall ), and other criteria using online search‑system feedback.

RLSF (Reinforcement Learning from Search‑System Feedback) : PPO optimizes the rewriting model with the QR‑RM reward. Training curves show steady reward increase and convergence.

Offline evaluation demonstrates significant improvements in both relevance and recall. Online A/B tests report a 4‑5% increase in gross merchandise volume (GMV).

Multilingual E‑Commerce Corpus and Continued Pre‑Training

Hundreds of billions of tokens are collected from internal logs, product catalogs, synthesized task data, and external e‑commerce webpages. After cleaning, the corpus is used for continued pre‑training (CPT) of the Qwen base model, yielding the ME‑LLM with enhanced multilingual e‑commerce knowledge.

eMMLU Benchmark

The eMMLU (e‑commerce Multilingual Multitask Language Understanding) benchmark v0.1 covers 30+ languages, 5 scenarios, and 19 tasks with tens of millions of examples. Evaluation shows ME‑LLM consistently outperforms the vanilla Qwen‑Base on most tasks.

Production Impact

ME‑LLM is deployed on platforms including AliExpress, Lazada, Daraz, Miravia, and TaoJP. In AliExpress and Lazada search, relevance scores improve by 5‑10 points, and query or title expansion driven by the model contributes a 4‑5% GMV uplift.