知识演化视角下论文与专利的热点技术方法对比分析——以人工智能自然语言处理领域为例

Comparative Analysis of Hot Technical Methods in Papers and Patents from the Perspective of Knowledge Evolution:A Case Study in the Field of Artificial Intelligence and Natural Language Processing

人工智能领域的研究成果在学术研究与工业实践之间转移流动。为揭示自然语言处理领域论文与专利中的技术方法的分布及其演化趋势,基于知识演化视角对比分析论文与专利技术方法的分布,将2013年1月到2022年12月arXiv数据集中自然语言处理领域的39025篇论文,以及按照“关键词+IPC分类号”的检索方法在incoPat专利数据库得到的14654件专利数作为样本,利用维基百科分类树构建方法本体,将论文和专利映射到领域本体中,并采用配对样本T检验方法验证两者知识演化方面的滞后性。结果表明:样本论文与专利存在共同关注的技术方法,包括卷积神经网络、循环神经网络、长短期记忆网络、注意力机制和预训练语言模型,其中前三者属于比较成熟的神经网络架构,而注意力机制随着上述神经网络架构的发展而迅速发展;专利技术方法相对论文滞后约1~2年。

The research outcomes in the field of artificial intelligence transition and circulate between academic research and industrial practice.To uncover the distribution and evolutionary trends of technical methods in natural language processing(NLP)within papers and patents,a comparative analysis is conducted based on the perspective of knowledge evolution.This paper analyzes 39025 NLP papers from the arXiv dataset spanning January 2013 to December 2022,and 14654 patents'information retrieved from the incoPat patent database by using the search method of"keywords+IPC classification number".Using the Wikipedia category tree,an ontology of methods is constructed to map the papers and patents to the domain ontology.A paired sample t-test is employed to verify the lag in knowledge evolution between the two sources.The findings indicate that the sampled papers and patents share common technical methods,including convolutional neural networks(CNN),recurrent neural networks(RNN),long short-term memory(LSTM),attention mechanisms,and pretrained language models.Among these,the first three represent relatively mature neural network architectures,while the attention mechanism has rapidly developed alongside these neural network architectures.Patent technical methods tend to lag behind papers by approximately 1 to 2 years.