Reproduce “TempRetriever: Fusion-based Temporal Dense Passage Retrieval for Time-Sensitive Questions” on OpenResearch

About this paper

Abdelrahman Abdallah, Bhawna Piryani, Jonas Wallat, Avishek Anand, Adam Jatowt

Researchers from the University of Innsbruck and Delft University of Technology developed TempRetriever, a dense passage retrieval model that explicitly integrates temporal information with semantic embeddings through various fusion techniques. The system consistently enhanced retrieval accuracy for time-sensitive questions and improved answer generation in Retrieval-Augmented Generation (RAG) pipelines, surpassing traditional dense retrieval baselines by providing more contextually relevant passages.

Key insights

• Explicitly fusing temporal embeddings with semantic representations significantly improves retrieval accuracy and ranking quality for time-sensitive queries, with Feature Stacking consistently performing best.
• Temporal negative sampling strategies, particularly selecting negative examples from different years or the same year as the positive passage, are more effective than random sampling in training robust temporal retrieval models.
• TempRetriever's approach can effectively enhance existing temporal retrieval models and significantly boost the performance of large language models in RAG pipelines for temporal question answering.

Summary by alphaXiv.