"Identifying Disruptive Technologies via Dynamic Hypergraph Modeling and Large Language Models"

"Identification of potential disruptive technologies is crucial for formulating science, technology, and innovation policies. Existing studies not only exhibit limitations in semantic analysis and multi-dimensional relationship modeling but also fail to incorporate external influences such as policy, market, and social factors. To address these limitations, a large language model (LLM) is employed to extract sentence-level technical directions from heterogeneous scientific texts for topic modeling with BERTopic. Subsequently, multiple co-occurring themes within individual documents are synthesized with themes extracted from their respective citing literature to construct hyperedges, thereby encapsulating the underlying mechanisms of knowledge recombination and evolutionary trajectories. Furthermore, tracking the evolution of dynamic hypergraphs across time slices facilitates the extraction of multidimensional signals, thereby enabling a more forward-looking and systematic detection of early-stage disruptive technology seeds. Finally, we leverage an LLM to architect a multi-agent system incorporating perspectives from technology, enterprises, policy, industry associations, and venture capital to simulate the inherent conflicts and potential coordination of evaluation criteria among different stakeholders, thereby filtering candidate technologies with the highest market development potential. We applied our framework to an empirical study on hydrogen energy, underscoring the method\u2019s scientific rigor and practical utility. Our approach provides an interpretable and scalable paradigm for pinpointing disruptive technologies transcending the limits of traditional bibliometrics."

识别潜在颠覆性技术，对于制定科技与创新政策至关重要。现有研究不仅在语义分析与多维关系建模方面存在局限，同时未能纳入政策、市场与社会等外部影响因素。为解决上述局限，本研究采用大语言模型（LLM）从异构科学文本中提取句子级技术方向，结合BERTopic开展主题建模。随后，将单篇文档内的多个共现主题，与其对应引用文献所提取的主题进行融合以构建超边，从而封装知识重组与演化轨迹的内在机制。此外，对不同时间切片下的动态超图演化过程进行追踪，有助于提取多维信号，从而能够更具前瞻性与系统性地识别早期颠覆性技术萌芽。最后，本研究借助大语言模型构建多智能体系统，融入技术、企业、政策、行业协会及风险投资等多维度视角，模拟不同利益相关方评估标准的内在冲突与潜在协调机制，进而筛选出市场发展潜力最优的候选技术。本研究将所提框架应用于氢能领域的实证研究，验证了该方法的科学严谨性与实用价值。本研究所提方法为突破传统文献计量学局限、精准识别颠覆性技术提供了一种可解释且可扩展的研究范式。