Intelligence Theories and Methods

617 Multisource Data Fusion and Emerging Frontier Technology Topic Detection Based on Graph Neural Networks Hot! Tian Xuecan, Wang Li DOI: 10.3772/j.issn.1000-0135.2026.05.001 Against the backdrop of rapid technological advancements and profound shifts in the innovation landscape, an efficient detection of emerging frontier technologies through fine-grained traceable semantic unit-level data fusion has a significant practical value. In this paper, a multisource data fusion and emerging frontier technology detection framework based on graph neural networks is proposed, refining the granularity of data fusion and technology detection from both semantic and structural feature perspectives. The framework constructs three indicator sets, novelty, growth, and attention, centered around the core characteristics of emerging technologies, embedding these indicator values as node attributes into technology cooccurrence networks to enrich the semantic features of single-source networks. Alternatively, by leveraging the strengths of graph attention networks in capturing structural features, it achieves feature-level multisource data fusion, enhancing both granularity and interpretability of the fusion process. Using the field of quantum sensing as an empirical case study, experimental results demonstrate that the proposed method enables an efficient multisource data fusion through node alignment operations, with correctly predicted entity pairs appearing within the top one to three positions of the result list on average. The refined fusion of multisource data not only improves the granularity of emerging frontier technology detection but also effectively reduces misjudgment risks, validating the framework’s potential to enhance the accuracy and interpretability of emerging frontier technology detection. 2026 Vol. 45 (5): 617-628 [Abstract] ( 17 ) HTML (156 KB)  PDF (3716 KB)  ( 8 )

629 Application of Information Analysis Methods in Evidence-Based Decision Making under the Digital Intelligence Integration Environment: Elements, Frameworks, and Optimization Strategies Hot! Xia Yikun, Ye Junling DOI: 10.3772/j.issn.1000-0135.2026.05.002 Information analysis methods have accelerated the transformation from traditional to intelligent paradigms in the context of digital intelligence integration. Evidence-based decision-making, a core component of digital-intelligence-driven systems, serves as a key domain in the application of information analysis methods. Grounded in the context of evidence-based decision making, in combination with systems theory thinking, this study adopted a systematic literature review approach to explore the elements, framework, and optimization of information analysis methods. It identified the characteristics and interrelationships among three core dimensions, namely, evidence sources, analysis methods, and decision-making scenarios, and constructed a five-dimensional application framework incorporating elements, patterns, functions, actors, and contexts. The results suggested that the effective application of information analysis methods was achieved through the coupling of three-dimensional elements in the “data-evidence-decision” chain, shaped by multi-party collaboration and complex contextual interweaving. Accordingly, this paper proposes strategies, including enhancing data governance capabilities, rationally applying artificial intelligence technology for empowerment, and facilitating the integrated use and cross-validation of diverse methods. This study aimed to reveal the structural features and evolutionary logic of the application of information analysis methods in evidence-based decision-making, enhance the understanding of their operational mechanisms and functions, and offer strategic recommendations for their effective use in the context of digital intelligence integration. 2026 Vol. 45 (5): 629-642 [Abstract] ( 23 ) HTML (132 KB)  PDF (2222 KB)  ( 9 )

643 Identification of Potentially Disruptive Technologies by Integrating Semantic and Citation Features Hot! Qin Hao, Zhao Yiming, Ma Yakun, Zhang Zhixin, Wang Qigang DOI: 10.3772/j.issn.1000-0135.2026.05.003 Disruptive technologies, as core drivers of industrial restructuring, play a critical role in shaping national technology strategies. To address the identification biases in existing research stemming from reliance on single-dimensional features, this study proposes a framework for identifying potentially disruptive technologies by integrating semantic and citation features. First, the PatentSBERTa model is employed to encode patent abstracts and extract semantic features, while a graph attention network is used to capture structural features from patent citation networks. Next, a dual-attention mechanism and a gated dynamic fusion module are designed and combined with a Bi-LSTM model to construct a supervised binary classification framework for identifying potentially disruptive technologies. The SHAP model is introduced to evaluate the contributions of semantic and citation features. Finally, an empirical study in the domain of carbon capture, utilization, and storage (CCUS) is conducted to demonstrate the application of the proposed framework, through which five potential disruptive technology directions and fifteen specific technology themes are identified. The results reveal that semantic features serve as the foundation for capturing the intrinsic innovativeness of technologies, whereas citation-based features provide critical external evidence for assessing their diffusion potential. Together, they form a complementary dual-verification mechanism of “intrinsic breakthrough-external diffusion.” In the context of innovation management under resource constraints, prioritizing technologies with intrinsic breakthroughs is recommended, followed by an assessment of their network niche based on citation features. This study offers a novel fusion analysis paradigm for identifying potentially disruptive technologies and provides actionable insights for national development policies and industrial planning in the CCUS domain. 2026 Vol. 45 (5): 643-662 [Abstract] ( 15 ) HTML (243 KB)  PDF (4470 KB)  ( 7 )

663 TreeGPT: An LLM-Based Automated Method for Constructing Technical Topic Architectures Hot! Wang Haofeng, Gao Yingfan, Wang Lijun, Yao Changqing DOI: 10.3772/j.issn.1000-0135.2026.05.004 The continuous growth of scientific and technological literature has posed significant challenges in the field of technical intelligence analysis, particularly in the efficient extraction of technical information from large volumes of unstructured text and the construction of a hierarchical semantic framework. This study proposes TreeGPT, an automated framework for constructing technical topic architectures using large language models (LLMs). By leveraging the intrinsic knowledge and semantic vectorization capabilities of LLMs, TreeGPT identifies technical topics and mines their relationships from scientific literature to generate a structured technical topic system for target domains. To validate the effectiveness of the proposed method, this study conducted a comparative empirical analysis against BERTopic and HLDA using the integrated circuit domain as a case study. Experimental results demonstrate that TreeGPT significantly outperforms traditional methods in terms of semantic accuracy and hierarchical clarity, while achieving an effective balance? between the performance and cost of LLMs. The proposed method also provides valuable support? for domain knowledge, semantic modeling, and technical intelligence analysis. 2026 Vol. 45 (5): 663-677 [Abstract] ( 19 ) HTML (148 KB)  PDF (4110 KB)  ( 9 )

678 Key Attributes and Indicators of GAI Training Data Quality from a Compliance Perspective Hot! Kuang Miaomiao, An Xiaomi, Lei Ming, Liu Hongyan DOI: 10.3772/j.issn.1000-0135.2026.05.005 To foster responsible artificial intelligence (AI), nations have introduced laws, regulations, policies, and standards that set forth explicit requirements for the governance and regulation of generative AI (GAI) training data quality. However, there remains a lack of technical specifications and guidelines for the implementation of these requirements, owing to the lack of studies on the key attributes and measurement indicators of GAI training data quality from a compliance perspective. Therefore, this study adopted a compliance perspective to analyze the requirements of the normative documents of laws, regulations, policies, and standards for training data quality. It identified four key attributes of GAI training data quality: accuracy, diversity, representativeness, and authenticity, and characterized them through the dual dimensions of the process and results. Following a three-round screening process by four researchers, an initial framework comprising 25 measurement indicators for the key attributes of the GAI training data was established. Subsequently, through a mixed-methods research approach, integrating field investigations, expert workshops, and questionnaire surveys, three iterative validation rounds were conducted. Ultimately, 20 indicators were validated, 5 were eliminated, and 6 were newly added, resulting in a final GAI training data quality measurement indicator system consisting of 26 core indicators. The proposed compliance-driven key attributes and measurement indicators provide a basis for measuring and verifying the quality of the GAI training data in conformity with existing laws, regulations, policies, and standards. This not only lowers the barriers and costs of compliance assessment but also offers an operable pathway for achieving the mutual recognition of data quality certifications and advancing the standardization of GAI governance systems. 2026 Vol. 45 (5): 678-688 [Abstract] ( 23 ) HTML (155 KB)  PDF (720 KB)  ( 4 )

Intelligence Technology and Application

689 Detecting Weak Signals of Frontier Technologies in Future Industries Through Outlier Processing Hot! Ye Guanghui, Tu Kai, Guo Lu DOI: 10.3772/j.issn.1000-0135.2026.05.006 The industry represents a concrete manifestation of transformations in productive forces. Amid the continuous emergence of technological innovations in future industries, such as artificial intelligence and quantum information, the prospective detection of frontier industrial technologies is essential for understanding the trajectory of scientific and technological development and supporting strategic planning. First, based on the theoretical premise that weak signals tend to emerge as outliers, this study constructed an outlier-feature indicator system for frontier technologies in future industries. From the perspective of signal amplification, the indicator system was integrated with an isolation forest model to filter technologically weak signals at the document level, and SHAP analysis was employed to validate the functional applicability of the model. Second, from a signal attenuation perspective, benchmark experiments involving models such as ChatGPT were designed to identify the optimal knowledge extraction model. By embedding relevant theoretical frameworks, including Hiltunen’s triadic model of future signs and Coffman’s weak signal research model, an intelligent weak signal detection model for frontier technologies in future industries was developed, supported by multiple theoretical underpinnings. Finally, from a signal visualization perspective, to address the issues of semantic deficiency and isolated interpretation, this study integrated approaches, such as social network analysis and topic modeling. Through multi-dimensional signal association, the identified weak technological signals were interpreted, and their evolutionary trajectories were anticipated. In the empirical analysis, the quantum information industry was selected as the domain for technology detection to verify the reliability, validity, and contextual applicability of both the outlier-feature indicator system and the proposed detection model. The results indicated that several technologies, including quantum state tomography, exhibit weak signal characteristics in the micro-level lexical signal dimension. In the macro-level thematic dimension, weak signals in areas such as quantum computing demonstrate stronger evolutionary intensity than those in fields such as quantum theory. These differentiated evolutionary patterns are closely associated with mechanisms such as industrial technological demand, commercialization potential, and disciplinary attributes. The outlier-feature indicator system and weak-signal detection model proposed in this paper not only provide a meaningful extension to existing theoretical frameworks in technology identification but also offer practical implications for proactive technological deployment and the allocation of scientific and technological resources by relevant stakeholders. 2026 Vol. 45 (5): 689-706 [Abstract] ( 15 ) HTML (241 KB)  PDF (7274 KB)  ( 8 )

707 Identifying Technological Evolution Paths and Predicting Innovation Opportunities from the Perspective of Technological Form Iteration Hot! Zhang Ke, Wang Congxiang, Wang Jinfeng, Zhou Wei, Feng Yicheng DOI: 10.3772/j.issn.1000-0135.2026.05.007 Understanding the trajectory of technological evolution is crucial for fostering innovation. However, existing studies often overlook the granularity of evolutionary themes and the multi-level structure of evolutionary relationships, which limits comprehensive analysis of the evolutionary patterns and constrains the efficiency and advancement of technological innovation. From the perspective of technological form iteration, this study proposes a systematic method for identifying technological evolution paths and predicting innovation opportunities. First, a patent citation network is constructed and hierarchically layered using an appropriate algorithm. A high knowledge persistence (KP) contribution-based patent traversal algorithm is then employed to extract the main evolutionary paths. Next, technical forms are identified from patents, and an initial technological form evolution network is established based on the keyword-citation-keyword (KCK) principle. To enhance network reliability, a triple-criteria screening framework—comprising semantic similarity, attribute consistency, and temporal continuity—is introduced to filter relationships within the network. Finally, the search path count (SPC) algorithm is applied to determine the principal technological evolution paths in the refined network. Building on the dimensional law coupling principle from multi-dimensional technology innovation map, the most recent technological forms along each main path are iteratively developed to identify and generate potential innovation opportunities. A case study on polypropylene modification technology is conducted to validate the feasibility and effectiveness of the proposed approach. The findings provide a valuable reference for uncovering fine-grained technological evolution paths and for accurately and efficiently identifying innovation opportunities. 2026 Vol. 45 (5): 707-723 [Abstract] ( 12 ) HTML (159 KB)  PDF (7239 KB)  ( 12 )

724 Technical Opportunity Discovery Based on “Problem-Solution”Statements Hot! Zhu Jianxin, Bai Wentao, Liu Ruinan, Lin Chaoran DOI: 10.3772/j.issn.1000-0135.2026.05.008 In the context of intensifying national technological competition, the ability to identify technological opportunities clearly and comprehensively is crucial for securing a strategic advantage. From the perspective of the theory of inventive problem solving (TRIZ), the core of technological opportunity discovery lies in the precise identification of technical problems and their corresponding solutions. However, existing approaches are constrained by limited global semantic analysis, the absence of systematic and iterative prompt optimization, and the inability of link prediction methods to fully capture and exploit complete semantic units. To address these limitations, this study proposes a novel framework for technological opportunity that ensures semantically explicit outputs and coherent, logically complete reasoning chains. First, the framework leverages the global semantic integration capabilities of large language models to extract technical object sentences, problem statements, and solution statements that are precise, generalizable, and semantically rich. Second, it introduces a prompt optimization strategy based on “multi-round iteration+multi-dimensional evaluation” to enhance the domain adaptability of prompts. Third, it integrates the TextRank algorithm with link relation statistics to identify high-value semantic units within technical solution statements. Specifically, TextRank is employed to filter salient information, while link relation statistics are incorporated to enable precise matching between problem statements and solution statements associated with technical objects. Concurrently, BERT, GraphSAGE, and the TPE algorithm are organically combined to ensure accurate prediction of link relationships. Experimental validation based on patent data from the aircraft engine domain demonstrates that the proposed framework effectively enhances semantic associations and strengthens the completeness of the logical chain in technological opportunity discovery. The findings provide valuable methodological support for innovators in identifying technological frontiers and guiding innovation practices. 2026 Vol. 45 (5): 724-739 [Abstract] ( 10 ) HTML (170 KB)  PDF (8382 KB)  ( 5 )

740 Differences in Knowledge Maturity Between Industry and Academia from the Perspective of Core-Periphery Structure Hot! Zhao Hongye, Zhao Yi, Zhang Chengzhi DOI: 10.3772/j.issn.1000-0135.2026.05.009 Knowledge maturity is crucial to academic innovation and industrial transformation. An accurate identification of differences in knowledge maturity between academia and industry can provide a basis for the allocation of innovation resources and talent development. However, existing studies often use entire papers as proxies for knowledge and estimate knowledge maturity based on the average publication year of references, overlooking the fine-grained knowledge embedded within papers and their heterogeneous maturation trajectories, thereby limiting the measurement precision. To address this, in this paper, we represent knowledge with fine-grained entities and propose a knowledge maturity evaluation framework based on the core-periphery structure, considering 1,845,637 papers in the field of artificial intelligence as research objects for the analysis. A computer science ontology classification model is used to extract knowledge entities from the papers and construct them as nodes in a dynamic knowledge network. Subsequently, a core-periphery partitioning algorithm is applied to identify the structural positions of these entities, trace their evolution path from the periphery to the core, and quantify their maturity using the interval between the first time when a knowledge entity enters the core and the publication time of the focus paper. Finally, the differences in the knowledge maturity utilization between academia and industry are examined, along with a multiple regression analysis. Entity evolution from the network edge to the core effectively characterizes maturity. This measure is significantly correlated with the paper novelty and disruptiveness. On this basis, the knowledge maturity of academic papers and academia-industry collaborative papers is confirmed to be significantly lower than that of industry papers. This study breaks through the previous analytical paradigm based on the overall literature and reveals such differences at the fine-grained knowledge entity level. 2026 Vol. 45 (5): 740-757 [Abstract] ( 12 ) HTML (251 KB)  PDF (2740 KB)  ( 7 )

758 Two-Stage Author Name Disambiguation Study Combining Rule-Based and Supervised Models Hot! Chen Yifan, Xie Ruixia, Yang Ning, Hu Wei, Zhang Zhiqiang DOI: 10.3772/j.issn.1000-0135.2026.05.010 Author name disambiguation (AND) is a foundational and critical task in academic research fields such as information retrieval, information integration, and bibliometrics. To address the limitations of single disambiguation models in satisfying practical requirements, this study proposes a two-stage automated author name disambiguation framework (TSD-RS) that combines rule-based and supervised models. In the first stage, the dynamic threshold method is employed to improve the rule-based model, thus enhancing preliminary disambiguation performance, while 12 rule-application orders are designed and compared for their effect on AND. In the second stage, a cluster network is constructed using paper clusters formed in the preliminary disambiguation as nodes and supervised model prediction results as edge weights. Subsequently, the InfoMap algorithm is applied for community detection to refine disambiguation iteratively. During this process, four automated training-set construction methods (for positive and negative sample pairs) and four supervised models (including large language models) are compared for their AND effectiveness. Experiments on three gold-standard datasets of varying scales show that the best disambiguation performance is achieved when selecting Order5 for rule sequence in TSD-RS’s first stage, the 1/2-shell method for positive sample extraction, and the random-forest model in the second stage. The resulting bF1 value attains a 95% confidence interval of 0.85±0.04, thus demonstrating improvement over baseline models. 2026 Vol. 45 (5): 758-775 [Abstract] ( 14 ) HTML (293 KB)  PDF (3868 KB)  ( 6 )