•  
  •  
 

Scientific Information Research

Keywords

patent claims; machine reading comprehension; GPT-4; Prompt; patent

Abstract

"[Purpose/significance]This study aims to automatically generate claims using the GPT-4 model, in order to reduce the writing difficulty for inventor and improve the work efficiency and quality. [Method/process]The article constructs Prompts suitable for automatically generating patent claims and implements four prompting strategies: ZeroShot, Exact-Drafting, Stepwise-Claim, and Exact-Step Claim. By inputting patent specifications and technical disclosure documents into the GPT-4 model and using Prompts to guide its output, the automated generation of patent claims is achieved. The ROUGE and BERTScore evaluation metrics were used to assess the quality of the text, and the generated text was analyzed in comparison with the reference text from multiple dimensions, including the number of claims, text length, high-frequency words, keywords, and common collocations. Finally, the quality of the generated claim documents was evaluated through expert assessment in five aspects: clarity, consistency, relevance, professionalism, and completeness. [Result/ conclusion]Empirical research shows that the Exact-Step Claim prompting strategy significantly improves the quality of generated claims; moreover, claims generated based on patent specifications are more closely matched in the number of claims and text length with the reference texts, indicating that the application effect of the GPT-4 model in the field of natural language understanding and generation is closely related to the quality of the input text. This study provides an efficient and intelligent assistance method that contributes to the development of the patent text writing and review field. However, there are challenges, and further improvements are needed for the model to precisely understand complex technical terms and comply with patent regulations, as well as to explore how to optimize the model's ability to judge the number of claims and the length of the text."

Digital Object Identifier (DOI)

10.19809/j.cnki.kjqbyj.2025.01.009

Download

Share

COinS