<p>As a significant achievement in the field of technological innovation, the analysis of patents holds an extraordinary research significance. Previous researchers have attempted to indirectly measure the technological innovation ability of patents by analyzing their novelty or impact based on citation and category information. However, such quantitative information is susceptible to human factors and cannot objectively evaluate the technological innovation of patents. In recent years, scholars have used various methods, such as keywords, topic models, and text similarity to analyze the textual information of US patents, but these methods require domain experts’ experience and knowledge to measure the innovation ability of patents. This study aims to use natural language processing (<i>NLP</i>) technology to analyze patent textual information and constructs a novel unsupervised learning framework for measuring the technological innovation of patents, including their novelty and impact indicators. In addition, we take the category of “nuclear physics; nuclear engineering” patents granted in China from 1993 to 2021 as an example to analyze and verify. The results demonstrate that the unsupervised learning framework can systematically and automatically measure the technological innovation of patents, and provide a new research method for analyzing patent quality.</p>