Researchers have developed a query-aware routing method to enhance cross-lingual retrieval performance in encoders. This approach uses a low-rank adapter, trained with frozen document embeddings, which is activated for cross-language queries while original encoders handle same-language queries. The SampoTron adapter, fine-tuned on the Nemotron-3-Embed-1B model, demonstrated a 20.9% relative gain in normalized discounted cumulative gain (nDCG) at rank ten on a financial benchmark, improving retrieval quality across English, Finnish, and Swedish directions. This method allows for specialized cross-language capabilities without altering existing document embedding vectors. AI
IMPACT Enhances cross-lingual information retrieval capabilities, potentially improving search and recommendation systems across different languages.
RANK_REASON Academic paper detailing a new method for improving cross-lingual retrieval in encoders. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
