MiniCOIL: Bridging Sparse Lexical and Semantic Retrieval
pub.towardsai.net·17h
Preview
Report Post

74 min read22 hours ago

Introduction

Search retrieval methods broadly fall into two streams: sparse lexical retrieval and dense semantic retrieval (see Fig. 1). Sparse methods, such as BM25, rely on exact keyword matches; they represent text as high-dimensional sparse vectors, where each dimension corresponds to a term, and only observed terms have non-zero weights. This yields precise, interpretable results but fails when the query and document use different words for the same concept. Dense methods, on the other hand, encode texts as low-dimensional dense vectors to capture semantic similarity beyond exact words. Dense retrieval can match synonyms or related concepts, but it sacrifices interpretability and often requires heavy neural infrastructure (ANN indexes, …

Similar Posts

Loading similar posts...