A Structured Generation Framework for Transforming Scientific Papers into Patent
The paper presents FlowPlan-G2P, a structured generation framework that converts scientific papers into patent descriptions. It decomposes the transformation into concept graph induction, section-level planning, and graph-conditioned generation. Experimental results show the approach outperforms standard models when evaluated on legal compliance criteria.
- ▪FlowPlan-G2P introduces a graph-mediated pipeline to translate scientific papers into patent language.
- ▪The pipeline consists of three stages: extracting a concept graph, partitioning it into section-aligned subgraphs, and generating text conditioned on these subgraphs.
- ▪Standard natural language generation metrics were found to favor legally non‑compliant outputs, prompting a domain‑specific evaluation.
- ▪Under this evaluation, FlowPlan‑G2P consistently surpasses vanilla proprietary models, highlighting the importance of structured decomposition over model size.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Computation and Language arXiv:2601.02589 (cs) [Submitted on 5 Jan 2026 (v1), last revised 23 May 2026 (this version, v4)] Title:FlowPlan-G2P: A Structured Generation Framework for Transforming Scientific Papers into Patent Descriptions Authors:Kris W Pan, Yongmin Yoo View a PDF of the paper titled FlowPlan-G2P: A Structured Generation Framework for Transforming Scientific Papers into Patent Descriptions, by Kris W Pan and 1 other authors View PDF HTML (experimental) Abstract:Generating patent descriptions from scientific papers is challenging due to fundamental rhetorical and structural disparities between the two genres.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
