Do large language models understand patents? Enhancing patent classification through AI-generated summaries

Abstract

Patent classification plays a crucial role in intellectual property management, but remains a challenging task due to the complexity of patent documents. This study explores a novel approach to enhance automatic patent classification by leveraging summaries generated by large language models (LLMs). Our approach involves using the GPT-3.5-turbo model to create concise summaries from different sections of patent texts, which are then used to fine-tune the RoBERTa and XLM-RoBERTa models for classification tasks. We conducted experiments on English and Japanese patent documents using two datasets: the well-established USPTO-70k and the newly developed JPO-70k, that we specifically created for this study.

Our findings show that models trained on AI-generated summaries – particularly those derived from patent claims or detailed descriptions – outperform models trained on original abstracts in both subclass-level multi-label classification and subgroup-level single-label classification. In particular, using detailed description summaries improved the micro-average F1 score for subclass-level classification by 2.9 points on the USPTO-70k and 3.0 points on the JPO-70k, compared to using original abstracts.

These results indicate that LLM-generated summaries effectively capture information relevant to patent classification from various sections of patent texts, offering a promising approach to enhance the accuracy and efficiency of patent classification across different languages.