Search papers, labs, and topics across Lattice.
This paper introduces TPOUR, a novel approach to unsupervised retrieval that optimizes for temporal relevance using Temporal Retrieval Preference Optimization (TRPO). By reinterpreting preference learning in the temporal dimension, TPOUR effectively guides retrievers to prioritize temporally aligned documents, addressing the challenge of retrieving semantically related but temporally misaligned content. Experimental results demonstrate that TPOUR significantly outperforms both unsupervised and supervised baselines, achieving notable improvements in nDCG@5 metrics despite being substantially smaller than competing models.
TPOUR achieves a 15.21% improvement in retrieval accuracy on implicit queries while being 72.7x smaller than its larger counterpart.
Unsupervised dense retrievers offer scalability by learning semantic similarity from unlabeled documents via contrastive learning, but they struggle to capture the temporal relevance, retrieving semantically related but temporally misaligned documents-an important aspect when a document collection spans multiple time periods (e.g., retrieving documents from 2018-2025 for "Who is the president in 2019?" introduces temporal ambiguity). Existing methods rely on supervised training with explicit timestamps, which are not always feasible. We propose TPOUR (Temporal Preference Optimization for Unsupervised Retriever), which uses our novel training method Temporal Retrieval Preference Optimization (TRPO). TRPO reinterprets preference learning in the temporal dimension, guiding the retriever to favor temporally aligned documents. TPOUR further generalizes to unseen time periods via interpolation in a learned time embedding, enabling continuous temporal alignment. Experiments on temporal information retrieval (T-IR), TPOUR outperforms both unsupervised and supervised baselines. Compared to Qwen-Embedding-8B, despite being about 72.7x smaller, TPOUR Contriever improves average nDCG@5 by +4.04 (+12.15%) on explicit and +4.98 (+15.21%) on implicit queries. We provide our code at https://github.com/agwaBom/TPOUR.
Microsoft Research