Pap2Pat: Towards Automated Paper-to-Patent Drafting using Chunk-based Outline-guided Generation
Knappich, Valentin, Razniewski, Simon, Hätty, Anna, Friedrich, Annemarie
–arXiv.org Artificial Intelligence
The patent domain is gaining attention in natural language processing research, offering practical applications in streamlining the patenting process and providing challenging benchmarks for large language models (LLMs). However, the generation of the description sections of patents, which constitute more than 90% of the patent document, has not been studied to date. We address this gap by introducing the task of outline-guided paper-to-patent generation, where an academic paper provides the technical specification of the invention and an outline conveys the desired patent structure. We present PAP2PAT, a new challenging benchmark of 1.8k patent-paper pairs with document outlines, collected using heuristics that reflect typical research lab practices. Our experiments with current open-weight LLMs and outline-guided chunk-based generation show that they can effectively use information from the paper but struggle with repetitions, likely due to the inherent repetitiveness of patent language. We release our data and code.
arXiv.org Artificial Intelligence
Oct-9-2024
- Country:
- North America
- Canada (0.04)
- United States
- New York (0.04)
- Minnesota > Hennepin County
- Minneapolis (0.04)
- Massachusetts > Middlesex County
- Cambridge (0.04)
- Mexico > Mexico City
- Mexico City (0.04)
- Europe
- Switzerland (0.04)
- Central Europe (0.04)
- Spain > Catalonia
- Barcelona Province > Barcelona (0.04)
- Germany > Saxony
- Dresden (0.04)
- Asia
- Middle East > UAE (0.04)
- China > Hong Kong (0.04)
- Japan > Honshū
- Kantō > Tokyo Metropolis Prefecture > Tokyo (0.14)
- North America
- Genre:
- Research Report > New Finding (0.93)
- Overview (0.93)
- Industry:
- Technology: