Scholarlink

<p>To support early-stage evaluation of patent applications, this paper proposes the A/X citation-guided grant prediction model (AXGPM), which combines a relational graph convolutional network with SciBERT embeddings to jointly perform examiner citation classification and patent grant prediction. Here, A-type citations refer to background prior art, while X-type citations signal strong conflicts that may invalidate novelty. By leveraging these citation types as early signals, AXGPM captures both the semantic content of patent texts and the structural information from citation networks. The model includes a citation-aware adjustment mechanism to enhance interpretability by modeling how different citation roles influence grant outcomes. Experimental results show that AXGPM consistently outperforms baseline models, including pre-trained small and large language models, and generalizes well across different time periods. We further validate the applicability of AXGPM through a science-to-technology case study, illustrating its potential to provide early insights into the transformation of scientific knowledge into patented technologies. These findings suggest that integrating citation-based graph learning with contextual language features offers a practical and interpretable solution for early patent evaluation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediction of patent grant using interpretable citation-guided graph neural networks

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Xinyu Tong,
  • Yonghe Lu,
  • Kazuyuki Motohashi

摘要

To support early-stage evaluation of patent applications, this paper proposes the A/X citation-guided grant prediction model (AXGPM), which combines a relational graph convolutional network with SciBERT embeddings to jointly perform examiner citation classification and patent grant prediction. Here, A-type citations refer to background prior art, while X-type citations signal strong conflicts that may invalidate novelty. By leveraging these citation types as early signals, AXGPM captures both the semantic content of patent texts and the structural information from citation networks. The model includes a citation-aware adjustment mechanism to enhance interpretability by modeling how different citation roles influence grant outcomes. Experimental results show that AXGPM consistently outperforms baseline models, including pre-trained small and large language models, and generalizes well across different time periods. We further validate the applicability of AXGPM through a science-to-technology case study, illustrating its potential to provide early insights into the transformation of scientific knowledge into patented technologies. These findings suggest that integrating citation-based graph learning with contextual language features offers a practical and interpretable solution for early patent evaluation.