Patent-CR: A Dataset for Patent Claim Revision (aclanthology.org)

Researchers have introduced Patent-CR, a groundbreaking dataset specifically designed for the patent claim revision task in English. This dataset comprises both initial patent applications that were rejected by examiners and their final, granted versions. Unlike standard text revision tasks, which typically focus on improving sentence quality, patent claim revision involves meeting complex legal criteria including clarity, technical accuracy, and legal robustness. This nuance underscores the dataset's significance as it opens new avenues for AI and machine learning applications in legal contexts. The study evaluates various large language models (LLMs), revealing that while GPT-4 outperformed its peers, many models produced ineffective edits that strayed from the required revisions. The research highlights promising outcomes from domain-specific models and emphasizes the value of fine-tuning techniques. Importantly, it also points out discrepancies between automated evaluations and human judgments, with the GPT-4-based assessments showing the strongest correlation to human evaluations. The Patent-CR dataset, coupled with these findings, lays the groundwork for future advancements in automated systems for patent claim revision, thereby enhancing the efficiency and accuracy of intellectual property processes.