Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Beyond Single Vector: Query-Aware Multi-Perspective Document Embeddings for Retrieval

AbstractRecently, large language models (LLMs) have demonstrated excellent performance on general information retrieval tasks, which encode query and document into an embedding vector representing semantic information respectively. However, single-vector representations often fail to capture the multifaceted semantics of informative documents, limiting the adaptability across diverse scenarios. To address this, we propose PRISM (multi-PeRspectIve Semantic inforMation retrieval), a novel framework that leverages LLMs’ implicit reasoning capabilities to extract and structure diverse semantic perspectives. PRISM introduces query-aware semantic fusion to dynamically integrate relevant perspectives during retrieval through weighted query embedding interactions. Furthermore, we propose the contrastive semantic diversification, a training mechanism that disperses semantic embeddings to maximize perspective diversity. Evaluated on the English retrieval subsets of massive multilingual text embedding benchmark (MMTEB), PRISM achieves state-of-the-art general results compared with models trained exclusively on public data.

More from EMNLP 2026

All EMNLP 2026 papers →