ARTFEED — Contemporary Art Intelligence

Sparse Coverage: A Semantic Retrieval Framework for Patent Prior-Art Search

ai-technology · 2026-08-19

A recent study introduces Sparse Coverage, an unsupervised framework for semantic retrieval aimed at patent prior-art retrieval, which overcomes the drawbacks of dense retrieval's reliance on single-vector representations. This method aligns local span embeddings with a sparse vocabulary of embedding-space centers chosen through a coverage-focused k-center objective. Testing on the CLEF-IP 2013 dataset reveals that Sparse Coverage achieves comparable or superior document-level recall to robust dense patent encoders, while also being effective for passage-level retrieval. Available on arXiv (arXiv:2608.16918), this framework does not depend on labeled training data, making it ideal for specific domains. Sparse representations facilitate efficient inverted-index searches, improving scalability and suggesting future potential in other technical document applications.

Key facts

  • Sparse Coverage is an unsupervised semantic retrieval framework for patent prior-art search.
  • It maps local span embeddings to a sparse vocabulary of embedding-space centers.
  • Centers are selected using a coverage-oriented k-center objective.
  • Spans activate nearby centers to produce sparse representations compatible with inverted-index retrieval.
  • Experiments were conducted on the CLEF-IP 2013 dataset.
  • Sparse Coverage matches or exceeds document-level recall of strong dense patent encoders in several configurations.
  • The method remains competitive for passage-level retrieval.
  • The paper is available on arXiv with identifier arXiv:2608.16918.

Entities

Institutions

  • CLEF-IP
  • arXiv

Sources