基于专利文本挖掘的细粒度技术机会分析

Research on Fine-Grained Technology Opportunity Analysis Based on Patent Text Mining

新一轮科技革命和产业变革中,技术机会分析在研发管理、企业决策中的战略地位不断攀升。然而,利用传统链路预测指标开展的技术机会分析精度已达到瓶颈,且固有的专家知识无法应对技术创新的动态性和复杂性,难以实现细粒度技术机会识别与分析。鉴于此,本文提出了基于专利文本挖掘的细粒度技术机会分析框架,该框架将专利文本挖掘和图神经网络链路预测法有机结合,将技术机会分析拆分为知识网络构建及演化分析、知识元素链路预测以及技术机会评估与筛选3个研究子任务。实证研究结果表明,利用多维关键词特征构建的知识网络能够完整呈现交叉领域的知识全貌,结合复杂网络指标和时间序列能进一步揭示技术发展脉络,为后续技术机会分析提供方向指引。BERT(bidirectional encoder representations from transformers)模型配合图神经网络方法适用于各技术生命周期的知识元素链路预测任务,相较于传统预测指标,BERT表现出更高的准确率和鲁棒性。经过与多源技术报告的对比评估,证实了基于该框架所析出的9个技术机会与计算机视觉技术发展实际情况相吻合,具备实际研发价值。

In the new round of scientific and technological revolution and industrial transformation,the strategic position of technology opportunity analysis in R&D management and corporate decision-making is growing.However,the accuracy of technology opportunity analysis based on traditional link prediction indicators has reached a bottleneck,the stubborn expertise can hardly cope with the dynamics and complexity of technological innovation,and fine-grained technical opportunity identification and analysis are difficult to realize.As a result,this study proposes a fine-grained technical opportunity analysis framework based on patent text mining that combines patent text mining and the graph neural network link prediction method and divides technology opportunity analysis into three research subtasks:knowledge network construction and evolution analysis,element link prediction and technology opportunity assessment,and screening.An empirical study in the field of computer vision shows that the knowledge network built using multi-dimensional keyword features can fully present the knowledge panorama of cross-fields,and the combination of complex network indicators and time series can further reveal the context of technological development and provide direction for subsequent technological opportunity analysis guidance.The BERT model combined with the graph neural network method is suitable for the knowledge element link prediction task of each technology life cycle,and it shows higher accuracy and robustness than traditional prediction indicators.Following a comparison and evaluation with multi-source technical reports,it is confirmed that the nine technical opportunities based on this framework are in line with the current development of computer vision technology and have practical R&D value.