基于大模型知识蒸馏的专利技术功效词自动抽取方法研究：以车联网V2X领域为例

Objective] This paper aims to improve the accuracy of automatic extraction of key technical words and corresponding function words from patent.[Methods] ChatGPT was used as the Teacher-model, and ChatGLM3 was used as the Student-model. Through knowledge distillation method, the training data generated by ChatGPT was used to fine-tune ChatGLM3, and multiple technical word extraction models and a function word extraction model were obtained. The technical words are extracted from the abstract, the first claim and the technical function paragraph, respectively, by using multiple technical word extraction models, and the function words are extracted from the technical function paragraph by using the function words extraction model.[Results] Compared with ChatGPT, the fine-tuned multiple technical word extraction models and function word extraction model show higher accuracy and lower recall rate, when extracting technical words and function words. The ChatGLM3 fine-tuning model of the first claim has the highest accuracy and F1 values of 0.734 and 0.724 respectively. Moreover, The accuracy of the function words extracted by the function word extraction model is 0.649, which is higher than the accuracy of the function words labeled by the commercial tool, which is is 0.53.[Limitations] The technical field and patent language of this research are single, the amount of patent verification data is small, and the data cleaning rules expect to be further optimized. [Conclusions] This research scheme improves the efficiency accuracy of automatic extraction of large language model through knowledge distillation operation. At the same time, this study can support the mining of cutting-edge innovative and hot technologies from patent texts, and support higher quality intelligent patent analysis.