Large Language Model Consensus and Corpus-Adaptive Topic Modeling for Classifying and Mapping Patent Innovation
Summary
- A design-oriented study proposes a patent analytics pipeline linking conceptual classification, topic formation, and interpretation.
- Three large language models separately classify patents by novelty and innovation type using a majority consensus mechanism.
- Topics are discovered by clustering PatentSBERTa embeddings, with parameter ranges scaled to corpus size and labels generated via LLM keyword analysis.
- Demonstrated on 30,964 public-sector-related patents, the pipeline produced 74 category-specific topic instances with a macro-average document coverage of 98.1%.
- Every model vote and piece of evidence is retained to ensure analysts can audit and revise outputs.
Why it matters
This auditable framework improves the transparency and reliability of LLM-assisted patent analysis, offering a reusable basis for technology intelligence and innovation mapping.