Improving Pinterest Search Relevance with Language Models: A Global Approach" (18 words)

Join us as we explore how Pinterest leverages large language models (LLMs) to improve search relevance, achieving significant performance gains across multiple languages and markets.

  • 1. Title: Integrating Language Model (LM) into Pinterest Search
  • 2. Presenters: Khan and Mukunda, machine learning engineers from Pinterest's search relevance team
  • 3. Pinterest is a visual discovery platform that helps users find inspiration for their lives
  • 4. Three main discovery surfaces on Pinterest: home feed, related pins, and search
  • 5. Monthly, Pinterest handles over six billion searches covering topics like recipes, home decor, travel, fashion, and more
  • 6. Pinterest search supports over 45 languages and reaches users in more than 100 countries
  • 7. Pinterest search backend comprises of query understanding, retrieval, ranking, and blending stages that produce relevant and engaging search feeds
  • 8. Focus on semantic relevance modeling at the ranking stage and how LM improves search relevance
  • 9. Presenting four key learnings from using LMs in Pinterest search relevance
  • 10. Lesson 1: LMs are good at relevance prediction, with improved performance as model size increases
  • 11. Baseline for LM is Search Sage, an in-house content and query embedding; 8 billion LAMA mastery model provides a 12% improvement over the multilingual BERT-based model and 20% over search stage em
  • 12. Lesson 2: Multilingual LMs generate captions and user actions that can be useful content annotations
  • 13. Text representation for each pin includes title, description, VM generated synthetic image caption, board titles, and queries leading to the highest engagement
  • 14. Oblation studies indicate enriching features is helpful for relevance prediction; notably, user action-based features improve performance
  • 15. Lesson 3: Knowledge distillation used to productionize LM model
  • 16. Fine-tune multilingual language models using human labeled data from specific segments and generic features that scale across various domains
  • 17. Train student model with semi-supervised learning, using the teacher model's five-scale soft scores for training data
  • 18. Online student model serves relevant search results while being scalable and cost-effective
  • 19. Lesson 4: Relevance tuning produces rich semantic representations in embeddings from large language models (LLMs)
  • 20. Pin and query embeddings can be used across Pinterest for content representation, benefiting related pins, home feed, and other surfaces
  • 21. Question: How were open-source LLMs chosen for fine-tuning?
  • Answer: Experiments with different language models determined the best performance; 8 billion LAMA mastery model provided the best results
  • 22. Question: Role of LLMs in distillation and cross-encoder architecture
  • Answer: LLMs are used to distill a student model for predicting search relevance; cross-encoder structure helps better capture interaction between query and pin
  • 23. Question: Evolution from previous search systems to the current LM architecture
  • Answer: New system improves upon existing features, especially with visual language models for expanding beyond limited markets for relevance data
  • 24. Question: Handling multimodality in the embedding model
  • Answer: Visual captions effectively capture image content; enriching features improve performance, and the same model is used for all languages with a multilingual LM.

Source: AI Engineer via YouTube

❓ What do you think? What are your thoughts on the ideas shared in this video? Feel free to share your thoughts in the comments!