<ul><li>Pinterest Search is a key platform for users to discover content aligned with their needs.</li><li>The article focuses on enhancing the search relevance model at Pinterest.</li><li>A 5-level guideline is used to measure relevance between queries and Pins.</li><li>The model architecture involves using a cross-encoder language model for classification.</li><li>Various text features like Pin titles, descriptions, and image captions are utilized for representation.</li><li>Knowledge distillation is employed to create a lightweight student relevance model from the teacher model.</li><li>Semi-supervised learning is utilized to train the student model on a large dataset with billions of rows.</li><li>Offline experiments demonstrate the effectiveness of different language models and text features.</li><li>Online A/B experiments show a significant improvement in search feed relevance with the new model.</li><li>The article concludes with plans for future work to enhance the Pinterest search relevance system.</li></ul>

Improving Pinterest Search Relevance Using Large Language Models

Discover more