KALE: Using a K-Sparse Projector for Lexical Expansion
Published in ACM SIGIR International Conference on Theory of Information Retrieval (ICTIR), 2023
Luís Borges, Bruno Martins, and Jamie Callan
KALE projects dense text representations into a sparse latent vocabulary that can be searched with an inverted index. Evaluated on MS MARCO, TREC Deep Learning, and BEIR datasets, the method shows how learned latent terms can complement lexical representations and improve first-stage retrieval with a modest additional computational cost.