Sparse Retrieval

Sparse retrieval aims to efficiently retrieve information by representing text as sparse vectors of keywords, enabling fast searching using inverted indexes. Current research focuses on improving the semantic relevance of these keyword representations, often leveraging large language models to learn better keyword expansions and incorporating techniques like mixture-of-experts for scalability. This approach offers a compelling alternative to computationally expensive dense retrieval methods, particularly for large-scale applications and scenarios requiring fast response times, and is actively being explored for various tasks including image and text retrieval, and question answering.