Answers are grounded in this paper and cite their sources.
Demonstrates effective high-value patent identification using HeteGAT-Pat, suggesting improved management in technology.
High-value patent identification is important for technological innovation management and industrial competitiveness. However, conventional approaches often have limited ability to capture both the semantic complexity of patent texts and the heterogeneous relationships among patent-related entities. To address this problem, we propose HeteGAT-Pat (Heterogeneous Graph Attention Network for Patents), a heterogeneous graph attention network for high-value patent identification. The model combines relation-aware multi-head attention, trainable delta word embeddings, and a multi-scale convolutional encoder to learn semantic and structural representations of patent documents. A heterogeneous graph is constructed to represent multiple node types and semantic relations, allowing the model to jointly encode local textual context and cross-document structural dependencies. In addition, a node feature fusion strategy and a neighbour-sampling-based normalization mechanism are introduced to reduce semantic sparsity and structural heterogeneity. Experiments on a private lithography patent dataset and an auxiliary public patent text classification dataset show that HeteGAT-Pat consistently outperforms representative baseline methods. On the private dataset, the proposed model achieves an F1-score of 94.90%. We further conduct a word-level interpretation-oriented analysis by combining term frequency–inverse document frequency (TF–IDF)-based pre-filtering with an in-graph knock-out procedure, thereby providing corpus -level and document-level evidence on words associated with high-value patent identification. These results indicate that HeteGAT-Pat is effective for high-value patent identification in the lithography domain and may also be useful for broader patent-related text analysis tasks.
No takes yet. Share an insight, caveat, or question.
Xia et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: