Temporal Preference Optimization for Unsupervised Retrieval

The article introduces TPOUR (Temporal Preference Optimization for Unsupervised Retriever), a novel approach for enhancing unsupervised dense retrieval systems by addressing temporal relevance through a method called Temporal Retrieval Preference Optimization (TRPO). TPOUR demonstrates superior performance over both unsupervised and supervised baselines in temporal information retrieval tasks, achieving a 12.15% improvement in nDCG@5 compared to Qwen-Embedding-8B, despite being 72.7x smaller. This development is significant for practitioners as it enables more accurate retrieval of temporally aligned documents without the need for supervised training with explicit timestamps, thus enhancing the capability of AI systems in handling time-sensitive queries.