Patent Value Evaluation and Intelligent Claim Generation via Incremental Pretraining of Large Language Models With Patent Texts

ABSTRACT

In order to improve the practicability of patent value evaluation and claim generation in intellectual property scenarios, this study proposes a structured and value‐driven framework. First, the incremental pretraining strategy of structure awareness is designed to help large‐scale language models learn the relationship among technical paragraphs, legal constraints and the hierarchical structure of claims. Second, a semantic and structured value evaluation model is constructed to predict the novelty, protection scope and stability of patents. Third, the controlled generation method of value perception and double‐task closed‐loop optimization mechanism is introduced to guide claim drafting and improve semantic expression. The results show that this method achieves a score of 0.837 in the semantic understanding task, which is significantly better than the score of 0.737 in the general baseline model. In the evaluation of patent value, the average performance of this method reaches 0.790, while the average absolute error is reduced to 0.127. For claim generation, the evaluation scores of bilingual evaluation understudy and recall‐oriented understudy for Gisting reached 0.44 and 0.47 respectively. The structural integrity and legal consistency index reached 0.89 and 0.85 respectively. After closed‐loop optimization, the evaluation performance is improved by 9.3% and the power generation quality is improved by 11.2%. Incremental pretraining of structure awareness and value‐driven generation strategy can effectively improve the reliability of patent evaluation and drafting auxiliary tasks.