Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Analyticup E-commerce Product Search Competition Technical Report from Team Tredence_AICOE

Rakshith R, Shubham Sharma, Mohammed Sameer Khan, Ankush Chopra

TL;DR

The study tackles multilingual e-commerce search relevance for two binary tasks, QC and QI, under limited language coverage and a model-size cap. It uses translation-based data augmentation to fill missing languages, combined with negative sampling and LoRA-based fine-tuning of Gemma-3 12B and Qwen-3 14B within 15B parameters. Key findings show that QC benefits most from translation-augmented data, while QI gains from a mix including minority-class synthesis, achieving 4th place with an average private F1 of 0.8857. The work highlights the value of high-quality multilingual augmentation and carefully designed negatives for robust cross-lingual relevance in e-commerce search, providing practical guidance for similar multilingual retrieval tasks.

Abstract

This study presents the multilingual e-commerce search system developed by the Tredence_AICOE team. The competition features two multilingual relevance tasks: Query-Category (QC) Relevance, which evaluates how well a user's search query aligns with a product category, and Query-Item (QI) Relevance, which measures the match between a multilingual search query and an individual product listing. To ensure full language coverage, we performed data augmentation by translating existing datasets into languages missing from the development set, enabling training across all target languages. We fine-tuned Gemma-3 12B and Qwen-2.5 14B model for both tasks using multiple strategies. The Gemma-3 12B (4-bit) model achieved the best QC performance using original and translated data, and the best QI performance using original, translated, and minority class data creation. These approaches secured 4th place on the final leaderboard, with an average F1-score of 0.8857 on the private test set.

Analyticup E-commerce Product Search Competition Technical Report from Team Tredence_AICOE

TL;DR

The study tackles multilingual e-commerce search relevance for two binary tasks, QC and QI, under limited language coverage and a model-size cap. It uses translation-based data augmentation to fill missing languages, combined with negative sampling and LoRA-based fine-tuning of Gemma-3 12B and Qwen-3 14B within 15B parameters. Key findings show that QC benefits most from translation-augmented data, while QI gains from a mix including minority-class synthesis, achieving 4th place with an average private F1 of 0.8857. The work highlights the value of high-quality multilingual augmentation and carefully designed negatives for robust cross-lingual relevance in e-commerce search, providing practical guidance for similar multilingual retrieval tasks.

Abstract

This study presents the multilingual e-commerce search system developed by the Tredence_AICOE team. The competition features two multilingual relevance tasks: Query-Category (QC) Relevance, which evaluates how well a user's search query aligns with a product category, and Query-Item (QI) Relevance, which measures the match between a multilingual search query and an individual product listing. To ensure full language coverage, we performed data augmentation by translating existing datasets into languages missing from the development set, enabling training across all target languages. We fine-tuned Gemma-3 12B and Qwen-2.5 14B model for both tasks using multiple strategies. The Gemma-3 12B (4-bit) model achieved the best QC performance using original and translated data, and the best QI performance using original, translated, and minority class data creation. These approaches secured 4th place on the final leaderboard, with an average F1-score of 0.8857 on the private test set.
Paper Structure (12 sections, 4 figures, 2 tables)

This paper contains 12 sections, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Dataset Statistics
  3. Query-Category Path
  4. Query-Item
  5. Data Preparation and Augmentation
  6. Query-Category Path
  7. Query-Item
  8. Experimentations
  9. Query-Category Path
  10. Query-Item
  11. Discussion
  12. Conclusion

Figures (4)

  • Figure 1: Language-wise distribution of positive and negative samples in the QC training dataset before data augmentation.
  • Figure 2: Language-wise distribution of positive and negative samples in the QI training dataset before data augmentation.
  • Figure 3: Language-wise distribution of positive and negative samples in the QC training dataset after data augmentation.
  • Figure 4: Language-wise distribution of positive and negative samples in the QI training dataset after data augmentation.