Publications
ワークショップ (国際) LoSparse:Low-dimensional Sparse Retrieval Model
Kohei Uno, Shogo Kisa, Yohei Iseki
SIGIR 2026 Workshop on eCommerce (ECOM26)
2026.7.24
Two-stage retrieval has become a standard paradigm in Neural Information Retrieval, where an initial retrieval stage efficiently selects candidate documents, followed by a more sophisticated re-ranking model. While dense retrieval methods leveraging pre-trained models have achieved strong performance, sparse retrieval has recently attracted attention due to its advantages in exact lexical matching, indexing efficiency, and interpretability. However, existing sparse retrieval models often generate extremely high-dimensional representations, as their embedding dimensionality typically corresponds to the tokenizer’s vocabulary size. This results in substantial computational and memory overhead during both training and inference. Moreover, such high-dimensional representations are difficult to integrate with partial matching techniques, such as the T-overlap query. In this work, we propose LoSparse: Low-dimensional Sparse Retrieval Model, which improves efficiency through low-dimensional sparse representations while enabling flexible retrieval control using the T-overlap query.
Paper :
LoSparse:Low-dimensional Sparse Retrieval Model
（外部サイト）