Purpose/significance This study aims to develop a novel altmetric indicator to measure the out-of-circle degree of academic papers on the WeChat platform，reveal their out-of-circle patterns and characteristics，and enrich the development and application scenarios of Chinese local altmetric data sources. Method/process We collect WeChat mention data of CSSCI EMD papers from 2021 to 2023. The out-of-circle index of relevant papers are calculated based on the improved Jaccard algorithm and large model fine-tuning method. Case analysis，descriptive statistical analysis and thematic analysis are comprehensively used to reveal distribution characteristics of the indicator and content characteristics of EMD papers with high out-of-circle index. Result/conclusion ①The WeChat mention coverage rate of EMD papers is high. Among WeChat scientific official accounts mentioning EMD papers，those with strong disciplinary attribute but weak academic attribute account for the highest proportion，and those with strong academic attribute and strong disciplinary attribute contribute the most WeChat mention count. ②The distribution of the out-of-circle index，AS score，and DS score shows a highly concentrated and extremely uneven pattern，concentrated in the small value range. There is a positive correlation between AS score and DS score. ③The themes of EMD papers with high out-of-circle index are highly consistent with social，political and current hotspots，and revolve around two main subjects： enterprises and the government. WeChat scientific official accounts show increasing attention to themes such as digitization，DID and total factor production continues to rise. Major political events played an important guiding role in the focus of WeChat scientific official accounts. Innovation/limitation The out-of-circle index provides new ideas for evaluating the cross-domain communication of academic papers. In the future，it is necessary to further improve the calculation method of the out-of-circle index and build a more comprehensive cross-disciplinary scientific communication evaluation system.

Purpose/significance As social media becomes an important channel for the communication of academic achievements，uncovering the communication characteristics and influence mechanisms of academic reporting can provide theoretical support and practical guidance for enhancing social impact evaluation systems. Method/process Based on Lasswell’s 5W communication model，this study integrates descriptive statistics，analysis of variance，and intergroup comparisons to analyze the content features，media types，and audience feedback of academic achievement communications on the Weibo platform，exploring how multiple factors jointly contribute to communication influence. Result/conclusion The communication influence of academic achievements on Weibo is primarily shaped by the combined effects of communication content，media，and audience. Reports concentrate on public-interest fields such as medical health and psychology. Official institutions and mainstream media constitute the core communication network，scientific self-media and academic publishers contribute more original content，while non-scientific self-media expand communication reach. Media tend to report academic achievements with broad impact，whereas audiences prefer topics closer to their own interests. Although researchers and journals are less frequently mentioned in reports，they are more likely to stimulate public interaction.

Purpose/significance Current research rarely analyzes the specific impacts of policies cited by academic papers from a fine-grained perspective. This study aims to provide policymakers with decision-making insights for enhancing policy influence and transformation effectiveness，while offering academia a new perspective for understanding policy knowledge diffusion，thereby fostering greater synergy between policy communication efficiency and knowledge innovation. Method/process Focusing on the academic dissemination of policies within policy–academia interactions，this paper takes artificial intelligence policies and their citing academic papers as samples，extracting over 4000 sentence-level citation chains containing policy knowledge units. Using citation analysis，the study examines the thematic evolution，diffusion characteristics，and influencing factors of knowledge flow differences of policy knowledge units from the perspective of knowledge diffusion. Result/conclusion The findings reveal that，thematically，policy topics such as AI education，AI ethics，and AI justice have attracted substantial scholarly attention and citations，making them key areas where policies exert academic influence. Analysis of knowledge flow differences shows that the dissemination power of policy knowledge units follows a long-tail distribution： a few units exert significant influence，while most have marginal impact. Key drivers influencing knowledge flow differences include policy topics and the administrative level of issuing agencies，while institutional co-authorship and the time lag of secondary citations also exhibit potential associations. From a temporal perspective，policies generally trigger academic responses in the short term （within three years），while some exhibit delayed resurgence in knowledge flow，often linked to national strategies or industrial development. It is suggested that future policy design should strive for thematic focus and institutional collaboration，foster innovative activation mechanisms，and adopt new approaches to policy communication.

Purpose/significance In the application of generative artificial intelligence within public libraries needs to be alert to the erosion of “technical efficiency priority” to “public equity”. It is manifested in ethical dilemmas，such as data bias may leading to the exclusion of vunlnerable groups，algorithmic black boxes exacerbating recommendation homogenization，and lack of public algorithm literacy leading to cognitive domestication. The essence is the imbalance between the technological logic and the public attributes of libraries. Method/process This study draws on Rawls’ “Difference Principle” and Sun stein’s “Push Theory”，constructs a human-machine collaborative governance framework of “Fairness Calibration-Compensatory Design-Collaborative Governance”： as the cire，and systematically responds to the imbalance of algorithmic justice from three dimensions：technology，system and culture. It provides theoretical references and practical paradigms for AI ethics governance in the public cultural domain. Result/conclusion Human-machine collaborative governance can not only maintain the efficiency of AI technology，but also achieve substantive fairness through a differential compensation mechanism. In the future，a cross-library data governance alliance needs to be established to solve the problem of data isolation，and the path for AIGC copyright compliance should be explored to promote the algorithmic justice of public libraries.

Purpose/significance With the rapid development of emerging technologies，large language models are gradually becoming an important channel for the public to obtain medical and health information and personalized health consulting services，however，when large language models generate natural language text，it is prone to contextual inconsistency，instruction inconsistency，logical inconsistency，etc.，i.e.，large model illusion，so it is of great significance to eliminate illusion and improve the accuracy of model-generated content. Method/process In this study，we take the open-source evaluation dataset “Huatuo26M-testdatasets” as the object of research，and propose a method to eliminate the illusion of medical Q&A based on cue word engineering and retrieval enhancement generation in order to reduce the degree of model illusion. On the one hand，the knowledge graph module is designed to use CPubMed-KG2.0 to obtain medical entity-related triples to update the knowledge of the large language model and enhance the model generalization ability； on the other hand，design a prompt strategy module that integrates the CRISPE prompt framework and chain-of-thought prompting to facilitate large language models in performing knowledge fusion on entities obtained from knowledge graphs and enhancing reasoning explanations. Result/conclusion Comparing the fluctuation of the hallucination rate before and after the hallucination elimination of the large language model，the experimental results show that the hallucination of the large language model is reduced，the overall hallucination rate of DeepSeek-V3 is reduced by 18.1%，the overall hallucination rate of Hunyuan-turbo is reduced by 12.4%，the overall hallucination rate of Doubao-1.5-pro-32k-250115 and QWQ-32B is reduced by 4.5% and 6.6%，as well as a decrease in the hallucination rate of subdivided hallucination types，providing an effective performance enhancement path for large language modeling in the medical Q&A domain.

Purpose/significance With the development of artificial intelligence technology，companion AI has become an important emotional media carrier in the digital age. Based on the theoretical context of the three-level theory of emotion，this paper systematically analyzes the complex interaction mechanism and emotional cognition landscape between users and companion AI，aiming to reveal the internal logic of user cognitive evolution and provide theoretical support for product optimization and industry development. Method/process Focusing on typical companion AI products in Apple App Store in China，the products are divided into two types according to function orientation： function retrieval class focusing on information service and role playing class focusing on emotional interaction. BERTopic topic modeling and ERNIE 3.0 sentiment analysis technology are adopted，combined with large language model-assisted semantic analysis method. From the three dimensions of instinct level，behavior level and reflection level，it systematically analyzes the user’s emotional needs，use experience and value cognition. Result/conclusion The research findings are as follows：①The initial emotion activation pattern was different in the instinctive level，and the key factors of initial impression were auditory，verbal and visual. ②At the behavioral level，there is a structural separation between rating and affective expression，and the relationship between affective engagement and affective deepening is shown as the interaction depth increases. ③The reflective level experiences cognitive expansion from individual value to social impact，contributing to rational criticism and boundary thinking of AI. Aiming at the problems of misalignment between technical ability and emotional needs，product logic and user experience in companion AI products，a multi-dimensional optimization strategy and development path are proposed.

Purpose/significance The instantaneous value of data elements，a key feature of the digital economy，requires a systematic mechanism to ensure efficient utilization. Method/process Grounded in the scenario-driven theory，this study proposes the “value flow theory，” analyzes the logic and features of instant data value，—combining the New York Citi Bike case—extracts a five-step pathway of identification，response，processing，transformation，and feedback，and finally comes out a scenario-driven release mechanism. Result/conclusion The study finds that instant data value is time-sensitive，scenario-dependent，network-diffusive，and nonlinear value added. Its mechanism is embodied in an operational structure of five interlinked stages forming a closed value loop，offering theoretical support for efficient value release. By introducing value flow theory and constructing a scenario-driven release mechanism，this research expands perspectives on time-sensitive data application and provides a practical framework for data

Purpose/significance A systematic and efficient strategic intelligence support system can provide decision-making support based on multi-source intelligence integration for government departments in strengthening the resilience of industrial chains and supply chains，as well as optimizing the layout of strategic emerging industries. Method/process This paper systematically analyzes the main risk points in current industrial chains and supply chains，sorts out the intelligence content and intelligence supply entities required to ensure the security of industrial chains and supply chains，proposes a collaborative model for multiple intelligence supply entities，and further develops an intelligence service model for government decision-making on industrial chain and supply chain security based on this collaborative model. Employing the integrated circuit industry as a validation case，this study demonstrates the feasibility and effectiveness of the proposed intelligence service model. Result/conclusion A systematic intelligence service model featuring the trinity of “goals-architecture-technology” is proposed，which includes the goal orientation of “formulating security decisions and ensuring the security of industrial chains and supply chains”，the service architecture of “intelligence generation，application，and evaluation”，and the auxiliary support of “artificial intelligence and blockchain technologies”. This model emphasizes the smooth collaboration among entities and the free flow of information，ensuring the precise and efficient supply of intelligence in a dynamic market environment，with a view to providing references for the construction of an intelligence service system for security decision-making in industrial chains and supply chains.

Purpose/significance In the era of digital economy，data elements，as a new type of production factor，have increasingly prominent strategic significance. Under the interaction of the policy innovation diffusion mechanism and the complex information environment，China’s data element policies exhibit the characteristics of a fragmented and scattered community. Systematically analyzing the community structure of data element policies and the distribution pattern of policy attention is not only a breakthrough point for resolving the bottleneck of market-oriented allocation reform of data elements but also provides theoretical support and practical guidance for enriching the data governance system of Chinese-style modernization. Method/process This paper follows the research logic of “theoretical framework-topic modeling-verification analysis”. By introducing the policy community theory，it explores the original community characteristics of China’s data element policies，constructs the data element policy community，and based on the structured topic model，captures and examines the distribution differences and evolution trends of policy sub-community attention. Result/conclusion The research finds that China’s data element policy community encompasses four sub-communities： guiding policies for data elements，combined and innovative concept policies in the data field，and cross-field integrated concept policies. There are significant differences in the distribution of policy attention among different sub-communities. Guiding policies and integrated concept policies pay more attention to the circulation and trading of data elements，while combined and innovative concept policies focus on data governance and openness and the construction of digital infrastructure. In terms of the time dimension，the attention of integrated concept policies to government service platforms and digital infrastructure construction fluctuates significantly，while the attention of the other three types of sub-communities to policy themes changes relatively steadily.

Purpose/significance As generative AI continues to evolve，the ability of prompt design has given rise to a new form of digital literacy in the intelligent digital era—prompt literacy. While a wide range of prompt guides have emerged，offering advice from various perspectives on how to craft effective prompts，a systematic analysis of these resources is still lacking. Method/process This study analyzed 44 representative prompt guides to identify their content and structural features. Thematic analysis was then employed to develop a framework for cultivating prompt literacy. Result/conclusion The analysis examined the features of prompt guides from three perspectives：language，target audience，and instructional approach. Additionally，this study proposes the preliminary RISE framework for cultivating prompt literacy，which encompasses four core dimensions： Recognition，Interaction，Shaping，and Ethics. This framework provides theoretical support for the future design of prompting strategies and the optimization of prompt engineering.

Purpose/significance In the quality-oriented scientific evaluation stage，exploring the advantages of the representative work evaluation system from a theoretical perspective and the essential differences from traditional quantitative evaluation indicators can help reveal the essence of this evaluation system and provide more effective suggestions for policy implementation and improvement. Method/process This article constructs a four-quadrant theoretical model of quality and quantity，dividing scholars into four types based on the quantity and quality of their achievements：“high quality and high quantity”，“high quality and low quantity”，“low quality and high quantity”，and “low quantity and low quality”. It analyzes the different orientations of representative work evaluation methods and traditional quantitative evaluation indicators such as publication quantity，h-index，total citation，and average citation in terms of quality and quantity. Based on MAG data，the theoretical model is empirically tested，using the number of published papers and the average impact factor of the journal where the paper is published to measure the quantity and quality. Result/conclusion Compared to traditional quantitative evaluation indicators，the representative work evaluation system can more effectively screen out “high-quality but low quantity” researchers，reflecting the quality-oriented function.

Purpose/significance It is of great significance to explore the use of information sources in managers’ environmental scanning under big data to support high-quality strategic decision-making to obtain enterprise competitive advantage. Method/process Firstly，based on the traditional information source theory，and integrating the theories of big data decision-making and new media，a new theoretical framework for information source classification is established，and then the influence mechanism model of the perceived quality，perceived availability and abundance of information sources on usage behavior is established. Then，using descriptive statistics and multiple regression methods，the survey samples were analyzed to reveal the new characteristics of information sources under big data，and the research hypothesis of key information source characteristics on scan frequency was verified. Result/conclusion It is found that the perception characteristics of four types of information sources： interpersonal-internal，interpersonal-external，non-interpersonal-internal，and non-interpersonal-external，show the characteristics of differentiated and complementary values. The new Internet information sources，represented by Internet media monitoring，has outstanding value in decision-making function. However，some traditional information sources are still the most trusted information sources for managers，and they have irreplaceable advantages. In addition，there is a significant positive correlation between the perceptual quality，perceptual availability，perceptual richness and scanning frequency of information sources. The correlation between perceived quality and scan frequency is the most significant. This study provides theoretical basis and practical guidance for the improvement of the scanning ability of enterprise managers under big data.

Purpose/significance This study explores the coping behavior process of college students in the context of intelligent recommendation. The findings contribute to optimizing algorithm design，promoting the healthy development of online platforms，and providing theoretical references for research on algorithm-related behaviors. Method/process Using grounded theory，a three-stage coding analysis was conducted on interview data from 25 college students，and a model of algorithmic coping behavior in intelligent recommendation scenarios was constructed. Result/conclusion The algorithmic coping behaviors of college students in the context of intelligent recommendation can be categorized into four types：algorithm acceptance，algorithm compromise，algorithm resistance，and algorithm avoidance. User needs，algorithm affordances，privacy concerns，recommendation quality，and awareness of the information cocoon trigger different coping behaviors through the mediating effects of emotional responses and cognitive evaluation. The assessment of coping behaviors directly influences subsequent coping choices，while user traits primarily play a moderating role. Moreover，the coping effect initiates a new cycle of behavioral responses，presenting a psychological trajectory of “acceptance – resistance – compromise – renewed resistance”.

Purpose/significance To understand users’ interaction with explainable artificial intelligence，compare the different feelings of users when facing the two explanation methods of providing thinking process and reference links，analyze the impact of different explanation methods and users’ familiarity with problem on their trust and quality perception of AI assistant，as well as the impact of users’ AI literacy on trust level. Method/process A controlled experiment was conducted to compare the effectiveness of two different explanation methods，DeepSeek R1 model providing a thinking process and Doubao large model providing reference links，in answering two questions of different user familiarity level. The trust level and quality perception of users towards AI assistants were obtained through scales such as HCTM and BUS 15. Result/conclusion ①When users are faced with unfamiliar problems，AI assistants providing either of the two explanation methods can significantly improve quality perception，and no statistically significant difference was found between them. ②When users are familiar with the problem，there is no significant difference in providing explanation or not. ③Users with high information literacy demonstrate higher levels of trust in AI assistants in various contexts. Innovation/value The research on innovation/value starts from the perspective of users and conducts empirical analysis on existing commercial grade AI assistants. The results are more practical and can provide reference for the design and optimization of AI assistants in the future.

Purpose/significance Accurately identifying the public’s sentiment contained in massive image-text multimodal online comment data is of great significance for deeply exploring the potential demands of the public and assisting the government and enterprises in scientific decision-making. Method/process Aiming at the problems of insufficient feature extraction and cross fusion，firstly，the paper introduced the CLIP multimodal large language model on the basis of BERT and ViT，and mapped image and text to the shared semantic space through comparative learning to bridge the semantic gap between image and text，and realize feature enhancement. Secondly，the paper constructed text-guided and image-guided cross-attention mechanism，and integrated with the self-attention mechanism as well as Fourier convolution to adequately learn the interdependence of features. Finally，the paper utilized global attention and residual connection to achieve the fusion of bottom-high level features to better improve the sentiment recognition accuracy. Result/conclusion The results of the comparison experiments，ablation experiments and case analyses demonstrate that the proposed model significantly outperforms other models in sentiment recognition. Besides，with the increase in the number of Epoch training rounds，the model recognition accuracy continues to improve，and could adjust to a relatively stable value quickly. These results indicate that the model has a better convergence effect and recognition performance. Innovation/limitation This paper proposes an image-text multimodal sentiment recognition model driven by large language model enhancement and multi-feature cross-fusion. In the future，more modalities such as video and audio would be incorporated into the research of sentiment recognition.

Purpose/significance To overcome the limitations of traditional bibliometric methods and provide theoretical support and methodological guidance for the early identification of scientific breakthrough papers. Method/process Based on the information expression system in rough set theory and the diversity measurement theory，this paper constructs a scientific breakthrough innovation identification model from a multi-indicator feature perspective，combined with machine learning classifiers for identification. Result/conclusion Using the annual breakthrough data from “Science” （2000–2024） for empirical validation，the random forest model achieved an accuracy of 0.9105，precision of 0.8624，recall of 0.9792，and an F1 score of 0.9171. Innovation/limitation We proposed adaptive measurement scheme for early identification of scientific breakthrough innovations based on rough set theory and diversity theory. However，only external features such as authors，institutions，funding，disciplines，and references of the papers were measured，without delving into content features.

Purpose/significance In this paper，we propose an innovative Hidden Markov Model （HMM） Multi Model Combination Framework to capture the dynamic characteristics of users’privacy emotion and its strong correlation with privacy behavior. Method/process More Than sixty thousands privacy-related comments on social media were subject extracted via BERTopic technology （Identify 5 core themes such as community supervision and personalized recommendation） and emotion intensity calculation. The research divides the evolution of privacy emotion into seven stages：initial anxiety stage，anger outbreak stage，fear spread stage，loss adjustment stage，trust establishment stage，emotion repetition stage and stable stage. The HMM Multi Model Combination Framework is used to analyze the probability of emotional state transition，clustering characteristics and behavior association patterns in each stage. Result/conclusion Empirical analysis shows that privacy emotional states （e.g. fear，anger） were significantly synchronized with high-risk privacy behaviors （e.g. disclosure of sensitive information），while trust states were associated with low-risk behaviors （e.g. anonymous browsing）；the pattern of emotional state driven behavior，such as "anxiety → fear" transitions often accompanied by high-risk behaviors；Negative emotions （especially high arousal anger and fear） significantly shorten the posting interval and accelerate information dissemination；Anxiety is the key hinge of emotional transfer，and “trust → loss”is the main path of trust collapse. Innovation/value Based on the emotion-behavior association model，this study proposes an active governance model strategy of “emotion–driven–path–prediction-layered intervention”. The results provide quantitative evidence and operational paths for social media platforms to upgrade privacy governance from passive response to active prediction.

Purpose/significance The core of sustainable and high-quality development of disciplines lies in knowledge spillover and innovation，and the impact of knowledge aggregation on knowledge flow is the key to the construction of disciplinary knowledge systems. Method/process Employing as data sources scholarly works from 23 humanities and social sciences disciplines during the 2014-2023 period，this study measured knowledge flow efficiency through the global SBM model. External indicators were selected based on agglomeration economy theory，with panel data regression and dynamic QCA methods employed to examine the mechanisms through which knowledge agglomeration factors influence knowledge flow efficiency. Result/conclusion The results show that related diversification agglomeration and competitive agglomeration positively impact knowledge flow efficiency，whereas specialization agglomeration and unrelated diversification agglomeration exhibit potential negative effects. Knowledge agglomeration characteristics demonstrate significant temporal and individual effects： social science disciplines should emphasize specialized and related diversified agglomeration，while humanities disciplines should focus on diversified and competitive agglomeration. Policy interventions could optimize configuration performance through knowledge layout adjustments. Given disciplinary heterogeneity，each field should coordinate influencing factors according to its unique knowledge structure and developmental needs to enhance knowledge flow efficiency.

Purpose/significance This study addresses challenges in depression detection on social media，including sparse depressive text features，cross-domain semantic gaps，and insufficient fine-grained analysis in existing methods. It aims to develop a fine-grained depression recognition framework integrating large language models and deep learning，applied to multi-dimensional depression identification for Weibo users. Method/process Structured prompt templates based on DSM-5 clinical diagnostic criteria were designed. The DeepSeek-v3 LLM was utilized to automatically extract symptom，emotional，and trigger features from heterogeneous user texts，constructing user-level symptom-emotion-trigger feature representations. A PLM-BiLSTM-Attention fusion model was further developed to perform depression severity grading and fine-grained trigger identification. Result/conclusion The “LLM + clinical diagnostic standards” approach enables semantic mapping between social media texts and clinical diagnostic criteria，achieving improved performance in feature extraction compared to traditional methods. The fusion model demonstrated good performance in both depression severity grading and trigger identification tasks. Experiments validated the framework’s practicality in the social media context，providing a feasible path for AI-based mental health screening. Limitations The current study is limited to Chinese unimodal data； future work can incorporate multimodal data to further enhance performance.

Purpose/significance To enhance the performance of target sentiment analysis tasks in the absence of large-scale，high-quality labeled data，this study explores the use of large language models for text data augmentation. Method/process A target sentiment analysis data augmentation method based on large language models is proposed，and three strategies are designed： direct rewriting strategy，sentiment change strategy，and multi-axis sentiment control strategy. These strategies generate new high-quality labeled data by rewriting the text，adjusting the sentiment polarity，and controlling the sentiment intensity，respectively. Result/conclusion The results based on the SemEval-14 dataset and the Deepseek large language model show that all three data augmentation strategies can improve the performance of the baseline model in the target sentiment analysis task. Mixing the original dataset with the augmented dataset achieves the best performance on the baseline model，verifying the quality of the augmented dataset and the effectiveness of the data augmentation method based on large language models in the target sentiment analysis task，which has certain theoretical and practical value. Innovation/limitation Three text data augmentation strategies based on large language models were proposed，effectively enhancing the performance of the target sentiment analysis task. However，the experiments were only conducted on one dataset and one large language model，and the experimental conclusions need to be further verified on multiple large language models and datasets.