【摘要】 [目的/意义]提出一种基于深度学习的前沿主题识别方法，使识别结果更加准确客观，为专利前沿主题的识别工作提供新思路。[方法/过程]采用专利之星检索系统数据库，通过SBERT预训练模型和聚类算法进行主题抽取工作；利用主题相似度计算和LDA模型对主题进行关联和标识；并从关注度和质量水平两方面构建前沿性指标，确定前沿主题。[局限]在数据源以及指标构建多样性方面尚需进一步研究和构建。[结果/结论]以我国制氢领域的专利数据作为实证研究对象，发现电解制氢、重整制氢、光制氢、水解制氢和燃料电池发电5个前沿主题，并通过对比分析验证了主题抽取与指标建立的正确性与有效性。更多还原

【Abstract】 [Purpose/Significance] This paper proposes a method for identifying frontier topics based on deep learning, which makes the recognition results more accurate and objective and provides new ideas for the recognition of patent frontier topics. [Methods/Processes] This paper uses CPRS as the database source. Firstly, topic extraction is based on SBERT pretraining model and clustering algorithm. Secondly, topic similarity calculation and LDA model is used to associate and identify topics. Finally, we construct the index of frontier topics identification from two aspects of attention and quality level. [Limitations] Further research is needed on the diversity of data sources and indicator building. [Results/Conclusions] Taking the patent data in the field of hydrogen production in China as the empirical research object, five frontier topics were found, including electrolytic hydrogen production, reforming hydrogen production, photo-hydrogen production, hydrolysis hydrogen production and fuel cell power generation. The correctness and effectiveness of topic extraction and index establishment are verified by comparative analysis.更多还原