| 摘 要: 随着专利数据规模的持续增长，传统专利质量评估方法在语义建模与结构关系刻画方面存在不足，现有方法在异构信息利用与语义关系建模方面仍有提升空间。为此，本文构建了面向专利质量评估的数据集，并提出一种基于异构图神经网络的评估模型(HG-HAN)。该模型融合PatentSBERTa提取的深层语义特征、结构化统计指标及时间特征，构建专利异构图，并通过改进的异构图Transformer架构刻画节点间复杂交互关系。同时，引入元路径权重自适应学习机制，以提升特征聚合效果。此外，结合“中国专利奖”权威标签与IPC同领域高引用专利构建基准数据集。实验结果表明，该模型在P、R及F1指标上均优于HAN、HGT等基线模型。 |
| 关键词: 专利质量评估 异构图神经网络 语义融合 |
|
中图分类号:
文献标识码:
|
|
| Patent Quality Evaluation Based on Heterogeneous Graph Neural Networks |
|
wufan
|
Nanjing University of Posts and Telecommunications
|
| Abstract: With the continuous growth of patent data, existing patent quality evaluation methods still face limitations in semantic modeling and structural relationship representation, and there remains room for improvement in exploiting heterogeneous information and modeling semantic relationships. To address these issues, this paper constructs a dataset for patent quality evaluation and proposesa heterogeneous graph neural network-based model, termed HG-HAN. The proposed model integrates deep semantic features extracted by PatentSBERTa, structured statistical indicators, and temporal features to construct a heterogeneous patent graph, and employs an enhanced heterogeneous graph Transformer architecture to capture complex interactions among nodes. Meanwhile, a meta-path weight adaptive learning mechanism is introduced to improve feature aggregation. In addition, a benchmark dataset is constructed by combining authoritative labels from the China Patent Award with highly cited patents within the same IPC categories. Experimental results demonstrate that the proposed model outperforms baseline models such as HAN and HGT in terms of Precision, Recall, and F1-score. |
| Keywords: Patent Quality Evaluation Heterogeneous Graph Neural Networks Semantic Fusion |
