 |
|
|
|
|
|
| Intelligence Theories and Methods |
|
|
925 |
Technology Genealogy Construction and Innovation Path Identification Based on Multi-granularity Semantic Computing Hot! |
|
 |
Yang Jinqing, Luo Xingyu, Xiong Bingqiao, Cao Gaohui |
|
|
DOI: 10.3772/j.issn.1000-0135.2026.07.001 |
|
|
As global technological competition intensifies and investments in research and development continue to increase, governments and investment institutions are increasingly pressured to identify novel paths that can drive technological innovation, mitigate technological uncertainty, avoid redundant efforts, and increase the output efficiency and social value of technological investments. Thus, unraveling the deep genealogical structures embedded within massive volumes of patents, clarifying the kinship among technologies, and tracing their inheritance chains are central to dispel the “technology fog” and detect “technological white spaces,” thereby facilitating the effective translation of patents and fostering industrial innovation. Building on the technological development theory, this study introduces the concept and framework of technological genealogy. First, we computed the patent similarity at the paragraph and phrase granularities. Subsequently, the optimal weighted combination of these granularities was determined by validating it against co-occurrence patterns in the international patent classification hierarchy, thereby quantifying the inheritance relationships among patents. A tree-structured technological genealogy was then constructed using patents as nodes and semantic associations as edges aligned along a temporal axis. Leveraging this genealogy, multiple types of innovation paths were identified, and the flow patterns of technological elements were revealed. Finally, an empirical analysis was conducted on 16899 patents for the B81B-007* technology in the 2.5-dimensional packaging domain. The results demonstrate that the technological genealogies exhibit complex dynamics of derivation and convergence while maintaining strong continuity. The discrete innovation paths focus on short-term advances, are highly diverse, and appear fragmented. The extended innovation paths capture the convergence and recombination of technological elements, demonstrating dynamic transitions from derivation to convergence. In addition, the sustained innovation paths span the entire timeline, forming coherent evolutionary paths, with later stages marked by iterative upgrades toward higher precision. |
|
|
2026 Vol. 45 (7): 925-939
[Abstract]
(
11
)
HTML
(223 KB)
PDF
(4024 KB)
(
4
) |
|
|
|
957 |
Research on the Underlying Logic, Practices, Methods and Implications of European Science and Technology Evaluation Hot! |
|
|
Jiang Tingting, Chen Yunwei |
|
|
DOI: 10.3772/j.issn.1000-0135.2026.07.003 |
|
|
Science and technology (S&T) evaluation, which plays an important role in realizing the “innovation-driven development” strategy, constitutes an indispensable component of S&T innovation systems. The substantial experience in S&T evaluations accumulated by European nations warrants systematic study and adaptation. Through a comprehensive literature review and comparative case analysis, this study systematically sorted the S&T evaluation frameworks of major European nations at both the macro- and micro-indicator levels. Their practical experiences in the innovation environment—culture, resources, and thinking—were synthesized while identifying regular patterns in evaluation metrics and methodologies. The study culminates in making four policy recommendations for an enhanced China’s S&T evaluation system: cultivate a healthy research ecosystem through standardized evaluation guidelines, develop innovative evaluation methodologies to align with emerging technological trends, align evaluation frameworks with national strategic objectives through legal institutionalization, and foster an open collaborative innovation culture to strengthen the S&T innovation momentum. The characteristic patterns and practical insights of European S&T evaluation regimes are further elucidated by providing substantive policy references for advancing China’s ongoing reform of scientific evaluation systems. |
|
|
2026 Vol. 45 (7): 957-968
[Abstract]
(
7
)
HTML
(121 KB)
PDF
(1220 KB)
(
1
) |
|
|
|
969 |
Factors and Their Relationships in Information Source Selection During Sudden Natural Disasters: Based on Cognitive Authority and Cognitive-Affective Trust Theory Hot! |
|
|
Li Yuelin, Zhang Xiangyihong |
|
|
DOI: 10.3772/j.issn.1000-0135.2026.07.004 |
|
|
This study explores the factors that influence information source selection in the context of sudden natural disasters by adopting a mixed research method for exploratory sequential design, to improve information disclosure quality and strategies. Based on the theories of cognitive authority and cognitive-affective trust theory, a literature review, and the results of focus group interviews, a theoretical model was developed for the factors influencing user information source selection in the context of sudden natural disasters. The model was verified through questionnaire surveys, structural equation modeling, and fuzzy set qualitative comparative analysis, exploring the conditional combinations of influencing factors in terms of different types of information sources. The results indicate that the quality attributes of information sources and cognitive trust have a significant impact on the willingness to choose official media, whereas the relationship attributes of information sources and affective trust have a significant impact on the willingness to choose nonofficial media. The willingness to choose interpersonal information sources is affected by all the above factors. Cognitive authority judgment plays a core role in the selection of different types of information sources. Task importance and urgency negatively moderate the impact of cognitive authority judgment on the selection of online information sources. The study revealed that the dynamics of cognitive authority judgments provides a new perspective for the application of cognitive authority theory in emergency situations, with implications for improving the quality of government information disclosure. |
|
|
2026 Vol. 45 (7): 969-985
[Abstract]
(
7
)
HTML
(250 KB)
PDF
(2698 KB)
(
2
) |
|
|
|
986 |
From Chaos to Spheres: Identification and Characterization of Social Risks in Short Videos of Emergencies Hot! |
|
|
Huang Shijing, Li Jiaxuan, Shen Hongzhou, Yuan Qinjian |
|
|
DOI: 10.3772/j.issn.1000-0135.2026.07.005 |
|
|
During emergencies, social risks in short videos are dispersed in form, structurally unclear, and rapidly evolving. Identifying their internal structure and evolutionary paths helps improve the precise governance of public opinion risks. In this study, we developed a “from chaos to clusters” analytical framework to identify and profile in short videos during emergencies. First, from the perspective of complex systems, we demonstrated the transformation mechanisms of short-video social risks from chaos to clusters and constructed a dual-axis space defined by the information-emotion and individual-collective dimensions, providing a theoretical basis for risk-cluster identification. Second, we established a multilevel feature engineering indicator system and combined a variational autoencoder (VAE), HDBSCAN (hierarchical density-based spatial clustering of applications with noise), and K-means to identify and profile risk clusters. Finally, based on Markov transition analysis, we examined the evolutionary paths and transition mechanisms of the risk clusters. The results showed that social risks in short videos during emergencies could be classified into six major risk clusters, which differ markedly in terms of risk intensity, content characteristics, and spatial distribution. Their evolution exhibited both escalation and divergence, with evolutionary paths varying across event types. By transforming dispersed risk information into structured social risk intelligence, this study enhances the understanding of the internal structure and evolutionary patterns of short-video social risks during emergencies and provides a decision-making basis for graded early warnings and differentiated governance of public opinion risks. |
|
|
2026 Vol. 45 (7): 986-1002
[Abstract]
(
7
)
HTML
(195 KB)
PDF
(3625 KB)
(
2
) |
|
|
| Intelligence Technology and Application |
|
|
1003 |
Research on Claim Feature-Driven Patent Pre-application Evaluation and Grant Probability Prediction Hot! |
|
|
Li Wang, Jia Zhixuan, Zhang Yiren, Cheng Fan, Ran Congjing |
|
|
DOI: 10.3772/j.issn.1000-0135.2026.07.006 |
|
|
A claim feature-driven model rooted in the “Three-Step Method” of patent examination was developed for patent pre-application evaluation and grant probability prediction, to address the critical imbalances between patent application volume and conversion rates in China, as well as the lack of quantitative mechanisms for pre-application assessment. The Levenshtein algorithm was first utilized in synergy with market transfer data to optimize sample selection by constructing positive and negative datasets that incorporate the dimension of “practical utility.” Subsequently, focusing on the core semantics of independent claims, claim dependency was introduced to quantify the scope of the technical features. This was integrated with a multilayer perceptron attention mechanism to achieve the dynamic weighted fusion of legal and semantic features, thereby realizing a consolidated representation of the “inventiveness” dimension. In constructing a cross-source heterogeneous prior art contrast pool and employing the K-means clustering algorithm, this study establishes adaptive dynamic evaluation thresholds to form a binary classification for the “novelty” dimension, ultimately achieving a pre-application assessment. Further embedding an autoencoder framework based on positive and negative sample comparisons allows the quantification of the evaluation results. Through the reconstruction error, the model characterizes the distribution consistency of the patents under evaluation within feature space, thereby refining the prediction of grant probabilities. Empirical results from the new-energy vehicle sector demonstrate that in the feature extraction stage, an F1-score of 0.919 was achieved by incorporating legal structural features into the model, significantly outperforming general semantic models. In the pre-application evaluation stage, the model identified a “recommended for application” rate of 25.28%. The model optimizes and supplements the existing grant standards by rigorously filtering low-market-value “dormant patents.” In the probability prediction stage, the average grant probabilities for the recommended and non-recommended groups were 72.50% and 47.50%, respectively, indicating significant discriminative power and statistical robustness. This research not only addresses the deficiencies of existing evaluation models regarding logical closed loops and legal feature utilization but also provides a scientific methodological path and technical support for universities and research institutes to improve patent application quality and promote efficient industry-university-research transformation. |
|
|
2026 Vol. 45 (7): 1003-1018
[Abstract]
(
6
)
HTML
(219 KB)
PDF
(2734 KB)
(
4
) |
|
|
|
1019 |
Emergency Decision-Making of Production Safety Accidents by Integrating Event Knowledge Graphs and Case-Based Reasoning Hot! |
|
|
Guo Yu, Liu Fangyu, Li Changfei, Zhang Haitao |
|
|
DOI: 10.3772/j.issn.1000-0135.2026.07.007 |
|
|
To improve the accuracy and intelligence of emergency decision-making for production safety accidents, this paper proposes an emergency decision-making method that integrates event knowledge graphs and case-based reasoning. This method addresses the entire process after an accident, including “element extraction, case matching, evolution reasoning, decision generation, and effect evaluation”. First, an event knowledge graph for production safety accidents is constructed based on multimodal data such as text, images, and videos. Through multimodal event extraction and relation recognition, the key entities, causal chains, and evolutionary logic of accidents are characterized, and a structured representation of accident knowledge is realized. Second, a case-based reasoning mechanism is introduced, and a case representation and retrieval method based on the event graph is designed. The path reasoning is used to identify high-confidence tasks and risk entities to realize the migration mapping from the historical disposal chains to the target accident situation and generate operable emergency decision-making suggestions. Finally, an empirical study is conducted for a typical emergency scenario of fire accidents, constructing source and target case bases to verify the effectiveness of case retrieval and consistency of decision generation. The unsupervised simple contrastive learning of sentence embeddings semantic consistency evaluation method is used for a quantitative comparison between the predicted and actual disposal schemes. The results indicate that the proposed method can effectively match similar cases and infer disposal tasks under conditions of incomplete information. In addition, the generated key decision recommendations encompassing command and dispatch, on-site control and risk isolation, casualty rescue, and personnel transfer are highly consistent with the actual disposal measures of the target cases. This study provides methodological support for rapid assessment, solution recommendations, and collaborative decision-making in emergency responses to production safety accidents. |
|
|
2026 Vol. 45 (7): 1019-1033
[Abstract]
(
5
)
HTML
(229 KB)
PDF
(2011 KB)
(
1
) |
|
|
|
1034 |
Hierarchical Discipline Classification with Multi-feature Synergy and Its Application to Interdisciplinary Measurement Hot! |
|
|
Zong Wei, Zhang Xin, Yang Xiaoxuan, Zhang Jianuo |
|
|
DOI: 10.3772/j.issn.1000-0135.2026.07.008 |
|
|
In this study, interdisciplinary measurements were conducted from the perspective of text content using hierarchical multi-label text classification methods to reveal the characteristics of interdisciplinarity more intuitively, thereby deepening and improving the interdisciplinary knowledge system. To address the problem of insufficient utilization of multidimensional features, such as paper text, the semantic and hierarchical structure of discipline labels, and the interactive relationship between paper text and discipline labels in the existing multi-label hierarchical classification of academic papers, a hierarchical multi-label discipline classification model based on multiple features, HMDCMF was constructed. The Centre for Research & Development Monitoring (ECOOM) discipline classification system was used to classify academic papers by discipline, and the interdisciplinarity of different disciplines was measured based on the discipline classification probability matrix. Compared with current mainstream hierarchical multi-label text classification models, the HMDCMF model performed the best. Interdisciplinarity was widespread across disciplines. Many disciplines are breaking through traditional development models, constantly integrating new fields and technologies and exploring new directions. By leveraging the HMDCMF model, the issues of insufficient semantic information in discipline labels and inadequate associations between paper text and discipline label features were alleviated, achieving interdisciplinary measurements from the perspective of paper content. At the theoretical level, this study expands the current methods of interdisciplinary measurement. In practice, it can promote the integration and collaborative innovation of knowledge among disciplines, providing a quantitative analysis basis for deepening the construction of an interdisciplinary knowledge system. |
|
|
2026 Vol. 45 (7): 1034-1050
[Abstract]
(
7
)
HTML
(243 KB)
PDF
(3110 KB)
(
1
) |
|
|
|
1051 |
Identification of Key Common Technologies for Future Industries Based on Augmented Neural Network Entity-Relationship Diagrams Hot! |
|
|
Hu Zewen, Xie Shaoke |
|
|
DOI: 10.3772/j.issn.1000-0135.2026.07.009 |
|
|
The accurate identification of key common technologies in future industries can provide model and intelligence supports for industrial technology prediction and management, policy-making, and the selection of research and development (R&D) and breakthrough directions for enterprises. In this study, we select the humanoid robot field as an example and construct an enhanced neural-network entity-relationship model that integrates a recurrent dilated convolutional neural network and an entity-relationship model. The five-tuple of technical entities and their semantic relationships from the field's patent documents are extracted, and a visual graph analysis is conducted. Then, the commonality of technologies is measured through indicators such as universality, benefit, and correlation, and common technologies are screened out based on the common technology score. Finally, the criticality of technologies is measured through indicators such as technological importance and leadership, thereby accurately and efficiently identifying key common technologies. The research results show that the enhanced neural-network entity-relationship model integrating a recurrent dilated convolutional neural network and five-tuple entity relationship model can effectively extract the five-tuple of technical entities and their semantic relationships from future industry patent documents. The results also reveal that the performance of the entity-recognition and relationship-extraction tasks improves, with the frequency 1 score increasing by 2.95 and 2.57 percentage points, respectively. The precision rate of the added entity-type prediction task in the model exceeds 94%. The constructed model, combined with key common technology-measurement indicators, can effectively identify the key common technologies in future industries covering eight key common technologies in the humanoid robot field such as motion control, joints, motors, heads, legs, sensors, positioning and navigation, and power systems. This model clarifies the technological breakthrough directions in the humanoid robot field. Its innovative achievements can drive the R&D of common technologies and industrial transformation, thereby upgrading the humanoid robot industry. |
|
|
2026 Vol. 45 (7): 1051-1067
[Abstract]
(
6
)
HTML
(206 KB)
PDF
(4241 KB)
(
1
) |
|
|
|
1068 |
Multiagent Collaborative Entity Linking for Historical Characters in Ancient Chinese Texts Hot! |
|
|
Yang Fan, He Jiacheng, Liu Chang, Zhang Qi, Liu Liu |
|
|
DOI: 10.3772/j.issn.1000-0135.2026.07.010 |
|
|
Historical figure entity linking in classical Chinese texts presents unique challenges that include diverse referential forms, absence of punctuation, and underdeveloped historical knowledge bases, rendering conventional entity linking methods inadequate for ancient literature processing. This study proposes a multiagent collaborative approach for historical figure entity linking in classical texts. Using the Records of the Grand Historian as the experimental corpus, a comprehensive historical figure knowledge base was constructed by integrating the China Biographical Database (CBDB) with manually annotated character-relationship data. Subsequently, we developed a collaborative framework based on the LangGraph multiagent architecture comprising three specialized agents: entity recognition, candidate entity construction, and candidate entity reranking. This method innovatively employs a dual-scoring mechanism that combines text similarity algorithms (such as BM25) with large language model (LLM) multidimensional semantic evaluation for entity mapping and ranking. Through inter-agent information exchange and negotiation mechanisms, the system achieves precise linking from textual entity mentions to standardized knowledge base entities. In experimental results, Qwen3-32B achieved optimal performance (recall@5: 0.908; accuracy: 0.947), representing improvements of 9.0% and 17.7%, respectively, over the bidirectional encoder representations from transformers (BERT) baseline. Ablation studies confirmed positive contributions from all modules, with the re-ranking mechanism showing the most significant impact, thereby validating that LLM semantic scoring is effectively enhanced by character background knowledge injection. This study presents a novel multiagent collaborative methodology for historical entity linking, offering new insights into entity linking research in the classical literature. |
|
|
2026 Vol. 45 (7): 1068-1082
[Abstract]
(
6
)
HTML
(169 KB)
PDF
(3809 KB)
(
2
) |
|
|
|
