Sparse Coverage matches dense patent retrieval with sparse centers

Patent prior-art retrieval is a recall-heavy task over long, highly structured documents. Dense retrieval improved semantic matching, but single-vector representations often compress multiple technical components, functions, and constraints into one embedding, losing detail. To fix this, researchers from Inria's ALMAnaCH, LISN, and Qatent (You Zuo, Kim Gerdes, Éric de la Clergerie, and Benoît Sagot) propose Sparse Coverage, an unsupervised semantic retrieval framework. Instead of one dense vector per document, Sparse Coverage maps local span embeddings to a sparse vocabulary of centers selected using a coverage-oriented k-center objective. Each span activates nearby centers, producing sparse representations that plug directly into inverted-index retrieval—the same infrastructure powering classic search engines.

On the CLEF-IP 2013 benchmark, Sparse Coverage matched or exceeded document-level recall of strong dense patent encoders in several configurations, while remaining competitive for passage-level retrieval. That's notable because it achieves high recall without training on labeled patent data, making it easy to deploy across domains. The sparse representations also enable fast, low-latency first-stage retrieval, ideal for filtering millions of patent documents before a more expensive reranker. Sparse Coverage is presented at CORIA-TALN 2026 and available on arXiv (2608.16918).