Dense retrieval and reranking for referenced provisions in electric power audit systems

Abstract

Electric power audits require practitioners to describe an audit issue and justify the final opinion by citing an appropriate referenced provision. In practice, the referenced provision should be retrieved from an authoritative provision corpus rather than generated, because correctness and traceability are critical in audit workflows. This paper proposes a dense retrieval and reranking framework for referenced provision retrieval in electric power audit systems. The method follows a two-stage pipeline: a two-tower dense retriever efficiently recalls a small candidate set (top-20) from a large provision corpus, and a one-tower scoring model performs fine-grained reranking by jointly modeling the audit problem description and each candidate provision. To strengthen semantic matching under audit-specific contexts, the audit issue category is incorporated into the reranking input. Experiments are conducted on a Chinese electric power audit text dataset, demonstrating that the proposed retrieval–reranking design provides an effective and practical solution for accurate referenced provision retrieval.