CPC-aware neural Patent retrieval with LLM-assisted query expansion

Abstract

Patent prior-art search demands high recall under severe terminological variation and multilingual noise, yet keyword-centric pipelines and coarse class filters frequently miss relevant documents. This study presents an integrated CPC-aware neural patent retrieval framework for Turkish aviation-focused patent abstracts, combining domain-adapted Transformer encoders with LLM-assisted query rewriting. We fine-tune PatentBERT for multi-label Cooperative Patent Classification (CPC) prediction at the section level ( C=6 : A, B, C, F, G, H) and use the resulting representations to support dense and hybrid ranking. To increase semantic coverage without deviating from the technical intent, user problem statements are expanded through controlled LLM-based paraphrase generation with dual constraints: embedding consistency with the original query and CPC consistency in predicted top sections. Optionally, an LLM-based normalization module can summarize and standardize verbose inputs to stabilize matching. Experiments on a labeled benchmark with graded relevance annotations show that the largest benefit of the LLM-assisted module occurs in the first-stage retrieval setting. Before reranking, CPC-consistent query expansion with retrieval-time normalization increases Recall@50 from 0.40 to 0.46, nDCG@50 from 0.33 to 0.36, and MAP from 0.21 to 0.24. After cross-encoder reranking, the additional gain is smaller, with Recall@50 increasing from 0.47 to 0.48, indicating that query expansion is best interpreted as a recall-oriented first-stage augmentation rather than as a replacement for strong reranking. In parallel, section-level CPC prediction improves by up to 6 percentage points in accuracy under the same evaluation protocol. The proposed modular design offers an auditable, scalable blueprint for multilingual patent engineering workflows, with clear separation between classification performance and retrieval quality.