Seismic: efficient and effective retrieval over learned sparse representation

Learned sparse representations form an attractive class of contextual embeddings for text retrieval thanks to their effectiveness and interpretability. Retrieval over sparse embeddings remains challenging due to the distributional differences between learned embeddings and term frequency-based lexical models of relevance, such as BM25. Recognizing this challenge, recent research trades off exactness for efficiency, moving to approximate retrieval systems. In this work1, we propose a novel organization of the inverted index that enables fast yet effective approximate retrieval over learned sparse embeddings. Our approach organizes inverted lists into geometrically-cohesive blocks, each equipped with a summary vector. During query processing, we quickly determine if a block must be evaluated using the summaries. Experiments on the Splade and E-Splade embeddings on the Ms Marco and NQ datasets show that our approach is up to 21× time faster than the winning (graph-based) submissions to the BigANN Challenge.