Multilingual Semantic Retrieval for Apple Music Search

2026-07-14 · Source: Apple Machine Learning Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, quick

Summary

Apple Music has deployed a new multilingual semantic retrieval system to enhance search recall across its 150+ storefronts and dozens of languages, particularly for challenging misspelled, transliterated, and cross-lingual tail queries. This system utilizes a 305M-parameter Siamese bi-encoder, fine-tuned from GTE-multilingual-base using curriculum-scheduled multi-objective training. Integrated via a hybrid retrieval architecture that combines dense nearest-neighbor results with the existing token-based index through quantile distribution matching, it avoids retraining downstream rankers. Offline evaluations showed a 69% relative improvement in Hit@10 over GTE-multilingual-base. A worldwide online A/B test demonstrated a 2.28% relative conversion-rate lift overall, an 86% reduction in the no-result rate, and significant gains for tail queries, which saw a 7.93% relative CR lift. This represents one of the largest search-quality improvements on the platform.

Key takeaway

For Machine Learning Engineers building multilingual search systems, this demonstrates that semantic retrieval offers substantial improvements, especially for tail queries. You should consider fine-tuning large bi-encoders like GTE-multilingual-base and integrating them using hybrid retrieval architectures with quantile distribution matching. This approach allows you to achieve significant conversion-rate lifts and reduce no-result rates without the extensive effort of retraining your entire downstream ranking stack.

Key insights

Multilingual semantic retrieval significantly boosts search recall for tail queries without impacting popular ones.

Principles

• Semantic retrieval enhances recall for complex, tail queries.
• Hybrid retrieval allows new model integration without full stack retraining.

Method

A 305M-parameter Siamese bi-encoder, fine-tuned from GTE-multilingual-base with curriculum-scheduled multi-objective training, is integrated via hybrid retrieval using quantile distribution matching.

In practice

• Deploy a 305M-parameter bi-encoder for multilingual search.
• Blend dense and token-based retrieval for robust search.
• Use quantile distribution matching for seamless model integration.

Topics

• Multilingual Search
• Semantic Retrieval
• Bi-encoder Models
• GTE-multilingual-base
• Hybrid Retrieval
• Quantile Distribution Matching

Best for: Research Scientist, AI Product Manager, Product Manager, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Simorgh at SemEval-2026 task 7: Region-Aware Hybrid Retrieval for Low-Resource Cultural Reasoning in Multilingual Question Answering — A region-aware hybrid retrieval system enhances multilingual QA by prioritizing culturally relevant evidence for efficient, stable inference.
• Not your usual search system — feat. BM25, FAISS and XGBoost — Production-grade search systems combine lexical and semantic retrieval with learning-to-rank for optimal performance.
• A Comparative Evaluation of Embeddings and LLMs in a Greek Book Publisher Setting - The CUP Dataset — Hybrid retrieval combining multilingual embeddings and lexical methods offers superior performance for low-resource language book search.
• Sifei at SemEval-2026 Task 8: Hybrid Retrieval and Query Rewriting for Multi-Turn RAG — The paper proposes a training-free hybrid RAG pipeline combining diverse retrieval and rewriting for multi-turn conversational contexts.
• SHIFT: Semantic Harmonization via Index-side Feature Transformation for Multilingual Information Retrieval — SHIFT corrects language bias in Multilingual Information Retrieval by adjusting document embeddings at indexing using relative language vectors.
• From Tokens to Concepts: Leveraging SAE for SPLADE — Replacing token vocabularies with SAE-learned semantic concepts improves sparse IR model efficiency and performance.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Apple Machine Learning Research.