Scholarlink

<p>Current studies on technology opportunity prediction often utilize link prediction on technical element networks, such as co-word networks, patent classification co-occurrence networks, and SAO-based networks. However, these methods typically compute network similarity based on topological structures to predict potential links. They overlook the embedded representations of technical elements and their relationships, which limits the ability to identify semantically related but topologically invisible connections. Moreover, these methods predict combinations of similar technical elements as technology opportunities, but rarely consider combinations of technical elements with multiple relationships as technology opportunities. Therefore, this paper proposes a method to better predict technology opportunities from the perspective of extracting and representing technical Subject-Action-Object (SAO) structures. Firstly, SAO structures are extracted from titles and abstracts of patents using dependency syntax analysis. Secondly, CompGCN model is used to learn entity and relation embeddings by aggregating multi-relational information from the SAO graph. Thirdly, ConvE is designed to capture heterogeneous feature interactions between entity and relation embeddings, and then infers new technical triplets as technology opportunities. The proposed method is validated in the field of speech recognition, the SAOs are extracted and provide a more detailed description of technology compared to keywords. In addition, SAOs that incorporate multiple relations outperform the single relations method with an improvement of 12.3% in MRR, and the method with a graph neural network representation outperforms that without representation learning with an improvement of 2% in MRR. Furthermore, the proposed method ConvE with CompGCN performs best in all prediction models, whose MRR reaches 23.1% and hit@5 reaches 33.6%. Finally, a case study is conducted to further validate the effectiveness of this method by utilizing Subject and Action to predict Object. These results prove that the proposed method is effective and useful. Simultaneously, this method is only a preliminary study and further research is needed in other areas with different multi-relation representation methods.</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

Technology opportunity prediction based on SAO representation learning

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Jinzhu Zhang,
  • Xuan Zhang,
  • Jialu Shi,
  • Mingxia Lu

摘要

Current studies on technology opportunity prediction often utilize link prediction on technical element networks, such as co-word networks, patent classification co-occurrence networks, and SAO-based networks. However, these methods typically compute network similarity based on topological structures to predict potential links. They overlook the embedded representations of technical elements and their relationships, which limits the ability to identify semantically related but topologically invisible connections. Moreover, these methods predict combinations of similar technical elements as technology opportunities, but rarely consider combinations of technical elements with multiple relationships as technology opportunities. Therefore, this paper proposes a method to better predict technology opportunities from the perspective of extracting and representing technical Subject-Action-Object (SAO) structures. Firstly, SAO structures are extracted from titles and abstracts of patents using dependency syntax analysis. Secondly, CompGCN model is used to learn entity and relation embeddings by aggregating multi-relational information from the SAO graph. Thirdly, ConvE is designed to capture heterogeneous feature interactions between entity and relation embeddings, and then infers new technical triplets as technology opportunities. The proposed method is validated in the field of speech recognition, the SAOs are extracted and provide a more detailed description of technology compared to keywords. In addition, SAOs that incorporate multiple relations outperform the single relations method with an improvement of 12.3% in MRR, and the method with a graph neural network representation outperforms that without representation learning with an improvement of 2% in MRR. Furthermore, the proposed method ConvE with CompGCN performs best in all prediction models, whose MRR reaches 23.1% and hit@5 reaches 33.6%. Finally, a case study is conducted to further validate the effectiveness of this method by utilizing Subject and Action to predict Object. These results prove that the proposed method is effective and useful. Simultaneously, this method is only a preliminary study and further research is needed in other areas with different multi-relation representation methods.