SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking

Dense retrievers lose the exact-match and interpretability properties of bag-of-words, while learned sparse retrievers had not matched their effectiveness.

A learned sparse first-stage ranker: an MLM-style encoder whose output layer produces term weights, trained end to end with EXPLICIT SPARSITY REGULARISATION and a log-saturation effect on weights, so representations are highly sparse and inherit inverted-index efficiency. Term weights are a learned version of BM25 expansion - they encode both importance and document frequency effects.

Competitive with state-of-the-art dense and sparse methods while trained end to end in a single stage; the paper also explicitly studies the effectiveness/efficiency trade-off by varying the sparsity-regularisation weight. Baselines named, but no absolute latency or hardware throughput figures.

Venue SIGIR 2021 short paper (ACM DOI 10.1145/3404835.3463098; arXiv comment '5 pages, SIGIR'21 short paper'). WHY IT MATTERS FOR INFERENCE COST: a sparse embedding is a few hundred non-zero term weights rather than a dense 768/1024-dim vector, so the first-stage scan moves far fewer bytes and can run on an inverted index or a CPU - the SPLADE-MLP variant does exactly that. It is also the mechanism behind the sparse head of M3-Embedding (pap-mm-bge-m3), which serves both retrieval functions from one checkpoint. First-stage-only use is the normal deployment: sparse retrieval then dense/late-interaction rerank. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. Peer-reviewed venue confirmed as SIGIR 2021 (same). This is an additive track annotation; `venue` is unchanged at 'other' and no status change is implied by the track alone. DOI CORRECTION 2026-10-10 (crawler/w11-doi): old DOI 10.1145/3404835.3463031 resolved to a different paper ('Cross-Graph Attention Enhanced Multi-Modal Correlation Learning for Fine-Grained Image-Text Retrieval'); corrected to 10.1145/3404835.3463098 which resolves to 'SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking'. VENUE RECLASSIFIED 2026-10-10 (crawler/w11-venue-enum): venue 'other' -> 'sigir', venue_kind 'conference'. Confirmed by record notes and agreeing publisher/index deposit in data/sources/.