TLDR: The DcuRAGONs team won the Alibaba International E-commerce Product Search Competition by developing a data-centric methodology. Their approach leveraged multilingual Large Language Models (LLMs), enhanced with translation augmentation and a two-stage task-adaptive pre-training process. They also introduced a robust cross-validation strategy to prevent data leakage. The Gemma-3-12B model, combined with these techniques, achieved top performance in predicting relevance between user queries and product items or categories across over 100 countries and 20 languages.
The world of e-commerce is vast and diverse, with millions of users searching for products across various languages and regions. Ensuring that a user’s search query accurately matches the right product or category is crucial for customer satisfaction and business success. This complex challenge was at the heart of the Multilingual E-commerce Search Competition, where the DcuRAGONs team from Dublin City University achieved the highest score with their innovative, data-centric approach.
The competition focused on two primary tasks: the Multilingual Query-Category Relevance Task, which assesses how well a user’s query aligns with a product category, and the Multilingual Query-Item Relevance Task, which evaluates the relevance of a query to a specific product listing. The dataset for this competition was massive, drawn from real-world search logs across Alibaba’s global platforms like AliExpress, Lazada, Daraz, Trendyol, and Mirivia, covering over 100 countries and 20 languages. This real-world data presented significant challenges, including noisy, ambiguous, and code-mixed queries, as well as a scarcity of expert annotations for many languages.
The DcuRAGONs Team’s Winning Strategy
The DcuRAGONs team, comprising Thang-Long Nguyen-Ho, Minh-Khoi Pham, and Hoang-Bao Le, built their methodology on the premise that a multilingual Large Language Model (LLM), specifically adapted for e-commerce, would be key to handling these noisy, diverse queries and generalizing to languages not seen during training. Their approach can be broken down into three core components:
- **Leveraging Multilingual LLMs with Translation Augmentation:** They systematically compared various small-to-large multilingual LLMs to find a balance between performance and efficiency. A crucial step was “translation augmentation,” where each original user query was also translated into English. This English translation acted as a “universal pivot,” helping the model understand queries from unseen or low-resource languages by grounding their meaning in a rich, high-resource representation, while the original query preserved fine-grained linguistic details.
- **Two-Stage Task-Adaptive Pre-training (TAPT):** Recognizing that general-purpose LLMs might lack e-commerce specific knowledge, the team introduced an intermediate pre-training stage. The model was first pre-trained on a related auxiliary task (like Query-Category or Query-Item relevance prediction) using a large-scale dataset. This “task-adaptive pre-training” helped the model learn domain-specific terminology, abbreviations, and semantic patterns prevalent in online shopping, significantly boosting its performance before fine-tuning on the final target tasks.
- **Category-Aware, Query-Grouped Cross-Validation (CA Split):** To ensure a robust and accurate assessment of their model’s ability to generalize, they developed a custom cross-validation strategy. For Query-Item tasks, all data points sharing the exact same query were grouped into a single fold, preventing the model from simply memorizing queries. For Query-Category tasks, categories sharing common prefixes (e.g., “Electronics > Audio Devices”) were grouped together. This forced the model to learn generalized relationships between queries and high-level product areas, rather than overfitting to specific leaf categories.
Model Architecture and Training
The team selected transformer-based multilingual LLMs as their backbone, specifically noting the strong performance of models like Gemma-3-12B, which supports up to 140 languages and fit within the competition’s parameter limits. To enhance training efficiency and stability, they utilized parameter-efficient fine-tuning via LoRA (Low-Rank Adaptation) adapters, allowing the model to specialize without losing its broad multilingual knowledge.
Experiments were conducted on powerful NVIDIA A100 GPUs, using the HuggingFace Transformers library with DeepSpeed for efficient distributed training. The models were fine-tuned for five epochs, optimizing with the AdamW optimizer.
Key Findings and Performance
Their experiments revealed several important insights. Larger multilingual LLMs consistently outperformed smaller architectures, demonstrating the need for robust representational capacity. Multilingual pre-training was critical, with models like the Gemma series showing clear advantages. While model size correlated positively with accuracy, the team observed diminishing returns, suggesting that architectural innovations or task-specific pre-training might be more impactful than simply scaling up models further.
The Task-Adaptive Pre-training (TAPT) stage proved to be a critical factor, consistently improving both Query-Category and Query-Item F1 scores. Even modest gains on the challenging private leaderboard reflected a meaningful enhancement, confirming that priming the model with domain-specific knowledge leads to better generalization.
Also Read:
- Boosting E-commerce Catalog Quality: A New LLM Prompting System
- OraPlan–SQL: Advancing Natural Language to SQL Conversion with Intelligent Planning
Looking Ahead
Despite their success, the DcuRAGONs team identified several areas for future improvement. These include exploring unified modeling approaches where a single model learns both Query-Category and Query-Item tasks jointly, investigating the impact of potential noise from low-quality translations, and explicitly optimizing models to exploit the hierarchical structure of category paths. Further research into sophisticated data augmentation techniques and task transferability could also lead to even more resilient and efficient search systems for real-world deployment.
The DcuRAGONs team’s work provides a valuable blueprint for tackling the complexities of multilingual e-commerce search, showcasing how careful data handling, translation-based augmentation, and parameter-efficient fine-tuning on powerful multilingual LLMs can lead to robust and high-performing solutions.
For more in-depth technical details, you can read the full research paper here.
