[Objective] This paper aims to enhance the utilization efficiency of governmental information resources and advance the intelligent transformation of public services by addressing the inherent knowledge limitations of general LLMs when processing policy texts. We investigate the effectiveness of a RAG framework to construct a more precise and reliable intelligent policy Q&A system. [Methods] This paper proposes a retrieval-augmented generation framework based on the Chinese policy large language model ChpoGPT. Specifically, the framework retrieves semantically similar policy documents from a knowledge base based on user queries and combines the retrieved results with ChpoGPT to enhance the model’s capabilities for downstream tasks. [Results] Experimental results demonstrate that our framework significantly outperforms existing models on key metrics. The ChpoGPT-based framework achieved a factuality score of nearly 90%. In terms of answer relevance, it scored 80.2%, outperforming the Gemini-1.0-pro model by 2.1%. Furthermore, it attained an answer semantic similarity score of 56.4%, surpassing the ERNIE 4.0 and Gemini-1.0-pro models by 4.1% and 2.8%, respectively. [Limitations] The language model still exhibits some uncontrollable behaviour in its answer output. [Conclusions] The retrieval-augmented generation of policy texts based on LLMs has certain reference value for the intelligent transformation of government services, but it still needs further improvement and optimization.

[Objective] This study investigates the performance differences among existing large language models (LLMs) in extracting entities and relations of Chinese medical text, and analyzes the influence of the number of examples and relation types on the extraction performance. [Methods] Based on prompt engineering approach, we use the API way to call 9 mainstream LLMs, modifying prompt from two perspectives: the number of examples and the number of relation types. Experiments are conducted using CMeIE-V2 dataset to compare extraction performance. [Results] (Ⅰ) The comprehensive extraction ability of GLM-4-0520 is in the first place, with F1 scores of 0.4422, 0.3869, and 0.3874 when extracting three relation types of “clinical manifestation”, “medication”, and “etiology” respectively. (Ⅱ) When varying the number of examples m in the prompt, the F1 score initially increases with m, and reaches a maximum score of 0.4742 when m=8, but it declines after m>8. (Ⅲ) After increasing the number of relation types to be extracted, n, the F1 score drops significantly: when n=2, the F1 score decreases by 0.1182 compared to n=1, and when n=10, the F1 score is only 0.2949. [Limitations] Currently, there are few public datasets available, so the experimental results are based on a single dataset. Additionally, since medical-domain LLMs are difficult to access via API, all models used in this study are from general domain. [Conclusions] The extraction performance varies greatly among different LLMs; A suitable number of examples can improve the extraction performance, but more is not always better; LLM is not good at extracting multiple relation types at the same time.

[Objective] This paper proposes a paradigm integrating large language models (LLMs) with knowledge graphs (KGs). We aim to address issues such as catastrophic forgetting, poor interpretability of generated content, and excessive demand for data and computational resources in vertical domain question-answering (QA) systems with fine-tuned LLMs. [Methods] First, we constructed a fine-grained KG for the traditional Chinese medical text “Treatise on Cold Damage”. Then, we employed a retrieval-augmented generation (RAG) model to incorporate this KG into a LLM through prompt learning to build a QA system. [Results] Compared to baseline models and fine-tuned models with professional data, the proposed system achieved a 14.67 and 1.33 percentage points higher satisfaction rate in subjective evaluations. In the objective evaluation, our model demonstrated an overall accuracy of 20.00 percentage points higher than the baseline models and 2.00 percentage points lower than the fine-tuned models. [Limitations] The application is limited to the traditional Chinese medicine domain related to the Treatise on Cold Damage. There is also a lack of standardized benchmarks to evaluate the system’s professional capabilities. [Conclusions] The proposed approach enhances the interpretability of generated content from vertical domain QA systems while substantially reducing the need for data and computational resources.

[Objective] In light of the fact that hate speech containing no obvious malicious words cannot be effectively identified in Chinese text, a Chinese hate speech detection method integrating multi-dimensional sentiment features (RMSF) was proposed. [Methods] Firstly, the RoBERTa model is used to extract both character- and sentence-level features from the input text, while sentiment dictionaries are used to derive multi-dimensional sentiment attributes. These character and sentiment features are then concatenated and fed into a BiLSTM network to capture deeper contextual semantic information. Subsequently, the output of the BiLSTM is concatenated with the sentence-level features derived from RoBERTa and processed through a multilayer perceptron before being classified using the SoftMax function. To address class imbalance, the focal loss function is applied during model optimization, thereby improving the accurate discrimination of hate speech. [Results] On the TOXICN dataset, the RMSF method achieves precision, recall, and F1 scores of 82.63%, 82.41%, and 82.45%, respectively. On the COLDataset, it achieves precision, recall, and F1 scores of 82.94%, 82.96%, and 82.85%, respectively. Compared to existing approaches, RMSF yields F1 score enhancements of 1.85% and 1.09% on the respective datasets. [Limitations] The hate speech detection method integrating multi-dimensional emotional features relies on tools such as sentiment lexicons. However, the extraction of emotional characteristics is constrained by the lexicon’s content coverage and semantic granularity. [Conclusions] The experimental findings indicate that incorporating multi-dimensional sentiment features into Chinese hate speech detection models can significantly enhance detection performance.

[Objective] Existing keyword extraction methods often suffer from limited attention scope, weak semantic representation, and restricted generative ability. To address these challenges, this paper proposes a patent keyword extraction approach (LLM-PKE) that integrates large language models with multi-feature networks. [Methods] LLM-PKE comprises three modules. In the extraction module, topic information is embedded into a Transformer attention network, combined with Graph Convolutional Networks to enhance sensitivity to topic terms and improve feature extraction. In the generative module, large language models produce keywords highly relevant to patent texts. In the ranking module, the large language model generates similarity scores for each keyword to remove synonyms and less relevant terms, yielding refined patent keywords. [Results] Compared to the best-performing baseline model, the proposed method improves the F1@5 metric by 1.98 percentage points. [Limitations] We use semantic similarity thresholds to remove redundant keywords; however, varying similarity standards across patent texts may limit accuracy and generalizability. [Conclusions] The LLM-PKE model outperforms existing approaches on patent datasets, offering a more effective solution for patent keyword extraction.

[Objective] This study proposes a dynamic topic modelling approach for short texts, driven by fine-tuned large language models (LLMs), aims to ensure high accuracy in topic identification and to reveal patterns in topic evolution. [Methods] The proposed method integrates instruction tuning, retrieval-augmented generation (RAG) and clustering techniques to improve the performance of topic identification. Topic evolution is modelled by establishing topic mapping relationships and conducting a time-sequential statistical analysis. [Results] Experiments on four short text datasets demonstrate that the proposed method outperforms the second-best baseline by an average of 6.15 and 7.71 percentage points in terms of topic coherence (TC) and topic diversity (TD), respectively. Ablation studies further evaluate the individual contributions of fine-tuning, RAG and clustering to the overall performance. Additionally, the method reveals distinct topic evolution patterns across datasets, including M-shaped and L-shaped trends. [Limitations] The current method does not use knowledge graphs to optimise the RAG component further, and its generalisability has yet to be validated across diverse, domain-specific short text corpora. [Conclusions] The proposed approach clearly outperforms others in terms of both topic identification accuracy and the ability to capture meaningful patterns of topic evolution.

[Objective] To address the insufficient attention in existing literature to the language style characteristics of rumors and the partially truthful dual-faced health information, this paper proposes a multimodal online health rumor detection model incorporating language style features (MWDLS: A Multimodal Wide and Deep Model for Online Health Rumor Detection Considering Language Style). [Methods] The MWDLS model leverages Aristotle’s rhetorical theory to extract persuasive language style features— appealing to emotion, logic, and character—and employs a bidirectional cross-modal interaction fusion strategy with a gating mechanism to achieve joint representation learning and classification prediction of shallow language style features and deep semantic features. [Results] We conducted extensive experiments on a real-world dataset from a leading Chinese social media platform and found that MWDLS outperformed the baseline models. It improved the F1 score of the target task by up to 11.98 percentage points. Notably, for the health rumor category and the dual-faced health information category, MWDLS increased the F1 scores by up to 16.63 and 11.71 percentage points, respectively. [Limitations] The current model does not examine other modalities, such as video and audio, nor does it incorporate large language models or knowledge-aware mechanisms to enhance early detection of health rumors. [Conclusions] By integrating language style features with multimodal deep semantic features, MWDLS effectively enhances the performance of online health rumor detection.

[Objective] This study aims to provide an efficient and accurate solution for multi-label classification of domain-specific scientific literature while offering insights into the potential applications of knowledge distillation techniques to other tasks. [Methods] We first used the large language models (LLMs) to generate preliminary labels for scientific literature in the target domain. Then, we employed knowledge distillation to transfer knowledge from the LLMs to a smaller model, enabling the training of a lightweight system for automatic multi-label classification. We evaluated the model on a tribology dataset by comparing the accuracy and F1 scores of distilled models with those of supervised models trained on human-annotated labels. [Results] The lightweight classification model trained with LLM-generated labels performs exceptionally well across multiple metrics, with accuracy and F1 score exceeding 0.96 and 0.86, respectively. [Limitations] The dataset diversity, the framework robustness, the label quality, and application scenarios require further expansion. [Conclusions] This research demonstrates the value of knowledge distillation for multi-label classification of scientific literature, highlighting its advantages in efficiency and cost-effectiveness. Future work may further refine the distillation process, expanding its applicability and supporting more tasks requiring efficient automated processing.

[Objective] To explore whether large language models (LLMs) and prompt engineering can replace classical deep learning models in the task of entity relation extraction from Chinese medical texts with high professionalism and domain characteristics. [Methods] This study uses three LLMs (GLM-4, ERNIE-4-Turbo, and DeepSeek-R1), and three classical deep learning models (CBLUE, CasRel, and GPLinker), to systematically compare the performance differences between LLMs based on prompt engineering and classical deep learning models. The comparison is conducted by varying the number of relation types to be extracted, the number of examples in the prompt for LLMs, and the training data size for classical deep learning models. We use BERT-Base and RoBERTa as encoders for classical deep learning models. [Results] Experimental results on the CMeIE-V2 dataset show that: (I) RoBERTa-CBLUE and RoBERTa-GPLinker achieve the best extraction results. When extracting one relation type, the F1 score reaches 0.5826 and 0.5853, and when extracting ten relation types, the F1 score is 0.5112 and 0.4934; (II) LLMs are not good at extracting multiple relation types simultaneously. When extracting two relation types, the F1 score of GLM-4, ERNIE-4-Turbo, and DeepSeek-R1 decrease by 0.1182, 0.0885, and 0.1310, respectively, compared to extracting one relation type; (III) adding examples to the prompt can improve the extraction performance of LLMs, but adding more examples does not necessarily lead to better results. [Limitations] This study is based on a single dataset, and future work could extend the experiments to datasets from other domains. [Conclusions] The prompt engineering approach for LLMs is currently difficult to replace classical deep learning models and can only be considered as an alternative when labeled samples are limited.

[Objective] This study aims to enhance the efficiency of policy information retrieval, enable intelligent analysis and comparison of policies, and provide precise decision support for policy formulation by constructing a structured policy knowledge base. [Methods] Using pro-business policies as a case study, we propose a framework based on large language models for efficiently comparing related policies. The framework consists of three core steps: knowledge base construction, retrieval and storage, and answer generation. [Results] Validation on datasets of pro-business policies demonstrates that the framework can automatically integrate multiple policies and perform semantic analysis to construct a knowledge base, supporting policy matching and comparative analysis. The Chroma-RAG model demonstrates clear advantages, achieving 60% on the Hit@1 index, 76% on the Hit@3 index, and 71.13% on the MRR index. Compared with traditional models such as TF-IDF, Word2Vec, USE, BERT, SBERT, DPR, and SimCSE, Chroma-RAG outperforms across retrieval metrics, underscoring the superiority of the proposed framework. [Limitations] The study primarily relies on cross-sectional data, which cannot capture the dynamic evolution of policies during implementation, thereby limiting deeper evaluation of policy impacts. [Conclusions] Knowledge base construction and policy comparison leveraging large language models significantly improve the intelligent analysis and comparison of policy texts. In particular, the approach offers strong decision-support capabilities in policy knowledge base development and comparative policy evaluation, providing valuable guidance for policymakers.

[Objective] This study aims to streamline the abstracting process of scientific literature and generate structured, knowledge graph-based summaries to facilitate efficient research. [Methods] Utilizing rice breeding publications from the PMC database, we constructed a dataset of 4,276 “full-text to abstract” Q&A pairs. By optimizing prompts and temperature coefficients, we used the Qwen2.5-7B-Instruct model for supervised fine-tuning. The fine-tuned model was integrated into the GraphRAG framework to generate graph-based abstracts for individual papers. Textual abstracts were generated via global queries using optimized prompts. [Results] The proposed method significantly outperformed the LightRAG baseline, with ROUGE-1, ROUGE-2, and ROUGE-L F1 scores increasing by 44.16, 61.36, and 54.87 percentage points, respectively. In human evaluations on a 5-point scale, the average improvement was 1.78 points, providing a more intuitive visualization of the logical relationships among key knowledge entities. [Limitations] Due to computational resource constraints, the study utilized a model with a relatively small parameter count, leaving room for further improvement in generative capacity. Additionally, the time-intensive nature of GraphRAG indexing suggests that inference efficiency requires further optimization for practical, large-scale applications. [Conclusions] Knowledge Graph-enhanced Retrieval-Augmented Generation effectively captures long-range implicit information within scientific texts. This approach produces more comprehensive, accurate, and hierarchically structured abstracts, thereby improving researchers’ reading efficiency and supporting scientific inquiry.

[Objective] This paper constructs a specialized framework for sharing and utilizing scientific data within the context of Artificial Intelligence (AI). [Methods] By systematically analyzing the data requirements of five representative AI tasks—traditional machine learning, large language model pre-training, large language model fine-tuning, Retrieval-Augmented Generation (RAG), and AI agents—we propose the FAIR×FAIR framework. This framework extends the traditional four FAIR dimensions (Findable, Accessible, Interoperable, and Reusable) by focusing on “For AI-Ready” principles and an associated layered technical stack. [Results] The FAIR×FAIR framework defines 13 technical requirements for AI-readiness, providing a systematic solution to bridge the semantic gap between AI tasks and scientific data. [Limitations] The practical efficacy of the proposed framework requires further validation through additional domain-specific application cases. [Conclusions] The FAIR×FAIR framework establishes a theoretical foundation and practical roadmap for scientific data sharing and efficient utilization in the AI era. It holds significant implications for advancing the evolution of data-driven research paradigms.

[Objective] In the process of generating summaries from textual and image information, when the multimodal contents are not entirely related to the reference summary, directly fusing them can introduce noise. To address this issue, this paper proposes a multimodal sentence summarization method based on large model theme enhancement. [Methods] We fine-tuned a large language model to produce high-quality theme and keyword information. Then, we used an attention mechanism to effectively fuse the theme with image information, reducing noise in the multimodal features. The original text was fused with the keywords to obtain a multimodal semantic supplementary feature with enhanced theme information. Finally, these two types of features are combined to generate the multimodal summary. [Results] On the public MMSS dataset, compared to the best-performing baseline Vision-GPLM model, our method improved ROUGE-1, ROUGE-2, and ROUGE-L scores by 2.79, 2.20, and 2.28 percentage points, respectively. [Limitations] The prompt templates utilized in large language model fine-tuning are relatively simple. We did not attempt fine-tuning with larger-parameter versions of the language model. The fine-tuning performance of the large language model can have a certain impact on the overall model performance. [Conclusions] By fine-tuning a large language model, this work reduces noise in multimodal features and enables different modalities to be fused while enhancing the model’s grasp of the main theme, thereby improving summary quality.

[Objective] This study reviews the current state of research on knowledge graph completion(KGC)technologies at home and abroad, aiming to provide a theoretical foundation for advancing the field. [Coverage] Using keywords such as “Knowledge Graph Completion”, “Link Prediction”, “知识图谱补全”, and “链接预测”, we retrieved and screened 133 representative papers from authoritative databases including Web of Science, Google Scholar, and China National Knowledge Infrastructure (CNKI). [Methods] Through literature review and synthesis, we evaluate existing research from three perspectives: model design, performance evaluation, and future directions. [Results] Relational semantics-based KGC models can be categorized into three types: those targeting complex relational semantics, those modeling connectivity-based relational semantics, and those addressing implicit, heterogeneous, and sparse relational semantics. On the WN18RR dataset, the SimKGC model—designed for sparse relational semantics—improves the MRR metric by 4.9%. On the FB15k-237 dataset, the DaBR model—focused on connectivity-based semantics—achieves a 3.4% improvement. [Limitations] Some emerging technologies lack large-scale benchmark validation. Due to the breadth of the field and volume of literature, not all relevant studies were covered. [Conclusions] Compared to traditional approaches, cutting-edge models demonstrate superior performance in knowledge graph completion. However, challenges remain in interpretability, scalability, multimodal and temporal data integration, and hallucination in large language models. These are the issues to be addressed by future research.

[Objective] To compare the performance of various reasoning methods based on Large Language Model (LLM) in policy question answering (Q&A) for public health emergencies. [Methods] Using DeepSeek-R1 as the base model, we implemented five reasoning settings: retrieval-augmented generation (RAG), knowledge graph (KG) collaboration, fine-tuning, web search, and reasoning without external data. All reasoning methods were evaluated against a manually annotated Q&A benchmark. For the setting without external data, we also included the performance of Qwen-QwQ and GPT-4o for comparison. [Results] The approach integrating RAG, KG collaboration, and fine-tuning achieved the best performance in the automatic evaluation, with BLEU-4 and ROUGE-L scores of 0.259 and 0.494, respectively. Web search ranked second, achieving BLEU-4 and ROUGE-L scores of 0.225 and 0.465, respectively. In the manual evaluation, the approach integrating RAG, KG collaboration, and fine-tuning achieved the highest content accuracy score (3.560). Large-scale models without external data performed better in terms of sentence fluency, answer completeness, answer usability, and answer credibility. [Limitations] The experimental data were derived from publicly available online policy texts and did not include private data from internal settings or multimedia formats. Therefore, reasoning performance in multimodal scenarios could not be verified. [Conclusions] With the support of local data, various reasoning methods can significantly mitigate hallucinations and improve the accuracy in policy Q&A, demonstrating their strong applicability to vertical domains such as public health emergencies. While external data improves the accuracy of LLM-generated content, self-contained reasoning without external data tends to yield more complete and credible results.

[Objective] This survey examines how retrieval granularity influences system performance in Retrieval‑Augmented Generation (RAG) and analyzes the trade‑offs and optimization strategies between contextual completeness and information precision across different granularities. [Coverage] Using the English and Chinese keywords “Retrieval-Augmented Generation” and “Retrieval Granularity”, we searched Google Scholar, ACM Digital Library, and CNKI for studies published between 2020 and 2025. A total of 106 representative papers were selected for in-depth analysis. [Methods] Centering on retrieval granularity, we conducted a comparative analysis of existing RAG approaches from the perspectives of technical pathways, core mechanisms, innovations, and limitations. [Results] We established a three‑tier framework of “coarse-grained, fine-grained, and hybrid-grained” retrieval. Coarse‑grained retrieval preserves contextual integrity but introduces noise; fine‑grained retrieval improves precision but risks semantic fragmentation; hybrid granularity aims to combine the strengths of both, but its fusion and scheduling mechanisms remain challenging. [Limitations] The review primarily focuses on text-based RAG methods, with less comprehensive coverage of multimodal RAG research involving images, audio, or video. [Conclusions] The advancement of RAG hinges on more intelligent granularity selection and refined information integration. Future research may explore proposition-level retrieval, dynamic granularity selection, adaptive mechanisms, and the coordination of structured and unstructured knowledge sources.

[Objective] The systematic utilization of policy documents is hindered by fragmented policy information, heterogeneous structural expressions, and the difficulties in capturing both explicit and implicit knowledge hinder. This study aims to address these challenges and strengthen top-level design and improve the efficient integration and sharing of policy information, by developing a generalizable construction method for policy knowledge graphs that can be adapted to domain-specific needs. [Methods] Drawing on a deconstructionist perspective, multiple theoretical frameworks are aligned to clarify the positioning of policy elements and to establish a multidimensional representation model as the schema layer. Fine-grained extraction techniques—integrating web retrieval, index matching, full-text parsing, entity recognition, and text classification—are designed to populate the data layer. Knowledge storage and visualization are implemented using the Neo4j graph database to generate the final knowledge graph. [Results] The proposed approach yields a universal ontology comprising 11 entity types, 14 relationship types, and 13 attribute types. Extraction standards for 29 fundamental elements and acquisition strategies for 9 domain-specific characteristic elements are defined. A set of efficient extraction methods and a dynamic visualization mechanism are developed. The feasibility and applicability of the method are demonstrated using 258 research integrity policies. [Limitations] As the construction process is primarily top-down, it may overlook innovative features embedded in grassroots policies, suggesting the need for future bottom-up enhancements to better capture emerging elements. [Conclusions] By integrating both general policy features and domain-specific characteristics, the proposed method supports comprehensive extraction of explicit and implicit policy knowledge, reveals deeper policy associations, meets diverse information needs across different fields, and strengthens cross-domain policy analysis and information integration.

[Objective] To address the challenges of efficiently fusing multi-modal information and inadequate modeling of correlations among emotion labels, this study proposes a multi-modal multi-label sentiment analysis method tailored for social media. [Methods] We constructed a model that integrates both modality-specific features and cross-modal shared features. Then, we utilized a cross-modal bridge connection mechanism to facilitate multi-modal fusion. Finally, we introduced a multi-head self-attention mechanism for multi-label prediction to capture the co-occurrence relationships among different emotion labels. [Results] Experiments on the CMU-MOSEI dataset demonstrated that the proposed model outperformed baseline models under various parameters and comparative settings. The ablation study validated the effectiveness of each module. Compared to the single-modality approaches using text, video, or audio alone, the proposed method improved accuracy by 11.4, 19.9, and 26.8 percentage points, respectively, indicating effective multi-modal fusion. [Limitations] In terms of system performance, the current approach cannot fully capture subtle emotional nuances. Furthermore, our dataset does not cover all possible emotional expressions and cultural contexts, highlighting the need for more diverse data. [Conclusions] The proposed model achieves effective modality fusion, yielding promising results in sentiment analysis.

[Objective] This study proposes a construction method for a cultural historical materials question-answering system that integrates knowledge graphs with retrieval-augmented generation (RAG), aiming to facilitate the dissemination and inheritance of museum artifacts and historical-cultural knowledge. [Methods] First, in light of the characteristics of cultural historical corpus texts containing numerous rare characters and classical Chinese, we developed a rare-character dictionary and a classical-modern Chinese parallel lexicon, and constructed a classical Chinese translation tool to reduce the hallucination and misinformation generated by large language model when processing historical documents. Second, we proposed a method based on prompt learning that integrates text vectorization similarity retrieval with knowledge graph retrieval, and designed task-specific prompt templates to achieve efficient knowledge reasoning and user interaction. Finally, we constructed the Qin Dynasty Cultural Intelligent Question-Answering System (ChatQDC) using Qin Dynasty cultural historical materials as a case study. [Results] Experimental results demonstrate that ChatQDC improves the average answer accuracy by approximately 8%, 20%, and 31% compared to GPT-4, ChatGLM3-6B, and DeepSeek-R1-8B, respectively. [Limitations] The current system is limited to text-based historical materials; its capability to process historical data in formats such as images and tables requires further enhancement. [Conclusions] Compared to traditional domain-specific fine-tuning approaches, the large language model in this system can be deployed without retraining, enabling cost-effective and energy-efficient vertical domain applications.

[Objective] U.S. congressional hearings generate vast amounts of text that are broad in scope and often expressed in colloquial language, posing challenges for intelligence analysis. This paper proposes a framework to automatically identify China’s technology security risks. [Methods] Starting from the features of hearings and the needs of analysts, we utilized large language models to integrate modules such as text filtering, summarization, and question-answering, thereby achieving high-quality intelligent identification. [Results] We examined our method with the 118th Congress hearing transcripts. The F1 score for text filtering, ROUGE-Lsum for summary generation, and the risk point recall rate for the QA system reached 0.7751, 0.6032, and 0.7636, respectively, significantly outperforming baseline models. [Limitations] Our method is primarily designed for U.S. Congressional hearing transcripts. Future work requires validation with more types of corpora to generalize and extend it into a universal approach. [Conclusions] The proposed method provides a powerful tool for extracting and analyzing technology-related intelligence from U.S. congressional hearings, offering valuable support for developing China’s technology security strategies.

[Objective] This study explores the usability and effectiveness of large language models (LLMs) in multimodal entity disambiguation tasks using chain-of-thought prompting. [Methods] We constructed a prompt template based on chain-of-thought reasoning for LLMs, integrating prior knowledge and multimodal information to help the model identify the correct referent entity from a set of candidates. [Results] On the Wiki-MEL, Twitter-MEL, and Weibo-MEL datasets, the proposed model achieved accuracy improvements of 15.1%, 11.5% and 4.1%, respectively, compared with the state-of-the-art models. [Limitations] The study did not thoroughly investigate performance variations in LLMs across different prompt construction strategies for the multimodal entity disambiguation task. [Conclusions] Chain-of-Thought prompting enables LLMs to better adapt to multimodal entity disambiguation tasks across diverse scenarios, demonstrating great application potential.

[Objective] This study proposes a deception risk identification model to address the insufficient multimodal feature representation in AI face-swapping fraud detection. [Methods] The FSFRI model captures the generation and dissemination characteristics of fraudulent information by extracting forged face video frame features, traffic description features, traffic payload features, and traffic temporal features. A feature fusion module is then employed to achieve cross-modal feature integration, followed by a risk identification module for assessing deception risks. [Results] On a simulated dataset, the FSFRI model demonstrated superior identification performance, achieving an F1 score of 0.920. It also exhibits strong robustness under low-noise conditions (noise ratio 0~0.2), with a minor F1-score decrease of 0.019 at a 0.2 noise level. [Limitations] The incorporation of multimodal features increases computational complexity, and the model’s performance under high-noise conditions requires further improvement. [Conclusions] By integrating multimodal features generated in AI face-swapping frauds, the proposed model effectively identifies deception risks. It provides valuable support and reference for intelligent prevention and control of telecom and online fraud.

[Objective] To enable early detection of rumors, this study investigates rumor detection models and methods based on text content. [Methods] A Chinese health rumor detection model using multi-scale graph neural networks enhanced by large language models is proposed. First, a text graph is constructed for each individual document to capture latent information in the sentence. Second, entity information is extracted from the text through prompt engineering to enhance knowledge representation. Finally, a multi-scale graph neural network with feature decomposition is employed to perform rumor detection. [Results] The proposed model achieves macro F1 scores of 95.21% and 87.39% on the CHECKED and LTCR datasets, respectively, outperforming existing baseline models. [Limitations] The proposed model relies solely on textual input and dose not incorporate multimodal data such as images or videos. [Conclusions] Utilizing large language models for knowledge enhancement not only facilitates efficient and accurate entity extraction but also enrich sentence semantic representations. The integration of multi-scale graph neural networks with feature decomposition enable effective capture of hierarchical features without compromising computational stability. Constructing individual text graph per document enhances flexibility and adaptability in downstream applications. By combining these components, the overall performance of the model is significantly improved.

[Objective] To support the pre-training, fine-tuning, and knowledge-enhanced reasoning of general and scientific large models, as well as the large-scale application of intelligent scientific literature information services. [Methods] This paper first analyzes the core challenges and intrinsic requirements faced by the AI4S paradigm, clarifying the conceptual connotation and key characteristics of the scientific literature knowledge base for AI4S. On this basis, it proposes a “data-model-service” three-layer theoretical framework for constructing this knowledge base, systematically outlines the key technologies and feasible implementation paths, and verifies the effectiveness of this theoretical framework through practical cases. [Results] Based on the aforementioned theories and methods, a preliminary infrastructure has been established, consisting of an AI-Ready data infrastructure centered on “Smart Data”, an intelligent model infrastructure centered on large language models for scientific literature and domain-specific models, and a multi-scenario-driven AI4S and AI4Data agent service infrastructure. This infrastructure has successfully supported research innovation activities in typical AI4S scenarios such as intelligent chemical engineering and digital cells. [Limitations] In large-scale practical applications involving multi-user collaboration and cross-domain scenarios, the theoretical framework of the scientific literature knowledge base for AI4S proposed in this paper still requires continuous verification and iterative optimization. [Conclusions] The constructed three-layer theoretical framework can provide a feasible pattern reference for scenarios such as the R&D of general and specialized intelligent models and the standardized processing of AI-ready corpora. The established scientific literature knowledge base will serve as a new type of research infrastructure to promote the improvement of pre-training efficiency and fine-tuning precision of general and scientific large models, as well as the enhancement of digital-intelligence capabilities including knowledge reasoning and computational analysis.

[Objective] This study explores the relationship between explainability, explanation accuracy, and information adoption behavior to optimize explainable designs for multi-scenario applications. [Methods] We conducted two parallel experiments in different contexts and examined the impact of AI explainability and explanation content relevance on users’ information adoption. Then, we explored the changing patterns in the AI “explanation singularity” across various task-specific information-demand scenarios. [Results] AI explainability significantly affects user information adoption, and its relationship is constrained by content relevance. Explanation singularities served as the turning point at which explainable designs influence information adoption behavior, and explanation singularities varied across different task contexts. [Limitations] We did not account for misinformation, and the exploration of the effects of XAI on user information adoption is limited. [Conclusions] The study reveals the mechanisms behind adoption for AIGC explanation from a multi-context perspective and uncovers the variation patterns of the “explanation singularities”. These findings highlight the pivotal roles of explanation relationships and content relevance in driving users’ information adoption, and emphasize their marginal impact on user behaviors.

[Objective] This study aims to exploit semantic associations across Chinese documents to enhance the performance of document-level event extraction. [Methods] We propose a Chinese document-level event extraction model (CSDEE) based on interactive semantic enhancement. The model utilizes an attention mechanism to construct a cross-document interactive semantic network that enhances entity recognition. The event extraction task is then completed through document encoding and event information decoding. [Results] The CSDEE model achieves a precision of 80.7%, a recall of 84.1%, and an F1 score of 82.3% in event extraction, outperforming existing baseline models. Ablation studies and generalization experiments on the ChFinAnn and DuEE-fin datasets further confirm the efficacy of CSDEE in Chinese document-level event extraction tasks. [Limitations] The current work focuses on improving document-level event extraction performance, without yet addressing multi-classification tasks involving overlapping event types. [Conclusions] Leveraging semantic similarities and associations across related documents can significantly improve the accuracy and robustness of document-level event extraction.

[Objective] This study focuses on the construction of the data enclave of the National Population Health Data Center (NPHDC), aiming to provide a more efficient, secure, and flexible environment for data processing and analysis. It addresses the challenges in centralized data storage, data security risks, limited computing resources, and urgent needs of users for data analysis and utilization. [Methods] The types, characteristics, implementation mechanisms, and scenario applicability of data enclaves were summarized. Combining the data application characteristics of NPHDC, we constructed its big data analytics platform based on the virtual data enclave approach integrating security enhancement, micro-segmentation, and artificial intelligence technologies. [Results] The big data analytics platform supports services such as data review, data processing, data analysis and mining, and peer review of publication-associated data for NPHDC. It has completed the review tasks of more than 32,000 datasets for more than 2,800 projects, more than 10,000 data analysis tasks, and more than 5,000 data processing tasks, with zero data leakage incidents and a resource utilization rate of 80%. [Limitations] The platform cannot achieve data sharing with decentralized storage across institutions. Further research needs to explore data enclaves based on privacy-preserving technologies such as multi-party secure computing and federated learning in combination with the development of NPHDC. [Conclusions] The platform effectively addresses the needs for secure sharing and collaborative analysis of population health data in a centralized manner, and is of great significance for the security protection and shared utilization of national population health scientific data.

[Objective] Existing literature on author name disambiguation lacks focus on Chinese-English collaborative author disambiguation, with one key reason being the absence of specialized and reliable datasets. This study addresses this issue by proposing a method to automatically construct an author disambiguation dataset using open internet resources. [Methods] Utilizing this method, we built a large labeled dataset, CHEN-AND, for Chinese-English collaborative author disambiguation research. Based on this dataset, we developed and evaluated several baseline disambiguation methods. [Results] The evaluation results show that the better-performing disambiguation methods achieved P-F1 and B3-F1 scores of 79.86% and 84.25%, respectively, which are significantly lower than the accuracy rates of mainstream English author disambiguation methods. [Limitations] CHEN-AND is focus on researchers in STEM fields, which shows biases from the actual disciplinary distribution of authors. This is mainly because one of the databases used for dataset construction, CSCD, is a literature database toward the Science, Technology, Engineering, and Mathematics (STEM) fields. [Conclusions] This study publicly releases the CHEN-AND dataset and disambiguation method evaluation results to facilitate future research on more efficient cross-lingual author disambiguation methods and the development of high-quality academic information exchange platforms.

[Objective] Cognitive diagnosis models (CDMs) rely heavily on prerequisite and similarity relationships among knowledge concepts, which pose challenges such as difficulty with automated modeling, sparse relational structures, and the absence of standardized evaluation criteria. To address these issues, this study proposes a multi-layer graph contrastive cognitive diagnosis model based on concept co-occurrence relationships. [Methods] First, we constructed a dual‑graph knowledge representation that incorporates both concept co‑occurrence and related co‑occurrence relations. Second, we employed multi-layer graph contrastive learning to iteratively enhance concept node embeddings. Finally, the enhanced embeddings were integrated with student-item interaction features and fed into diagnostic functions to generate cognitive diagnosis results. [Results] Experiments on three real-world educational datasets (ASSISTments09, MAT2016, and EdNet-1) demonstrate that the proposed model significantly improves diagnostic performance. Using ASSISTments09 and EdNet-1 as examples, the model achieves peak AUC values of 80.14% and 73.74%, respectively, outperforming the best baseline models by 1.64% and 1.32% respectively. Furthermore, the model is more interpretable than the baselines. [Limitations] In scenarios where items involve simple structures or assess only a single knowledge concept, the advantages of modeling co‑occurrence relationships may not be fully realized. [Conclusions] By integrating concept co‑occurrence graphs with an adaptive contrastive enhancement mechanism, the proposed multi-layer graph contrastive CDM effectively captures complex inter‑concept relationships. Experimental results confirm that the model substantially improves the accuracy of knowledge state prediction.

[Objective] Aiming to meet the demand for automatic acquisition of innovative content in scientific papers, this paper proposes a method based on enriched segment detection. [Methods] We identified innovation sentences in two stages. First, we constructed a keyword list for the enriched sections of innovative sentences and employed a sliding window scoring approach to locate these targets, thereby narrowing the scope of sentence identification. Then, we designed a Context-BERT model that integrates contextual information to identify innovation sentences automatically. [Results] The proposed approach achieved an F1 score of 87.27% on the test dataset, demonstrating effective and accurate identification of innovation sentences in scientific papers. [Limitations] The dataset used in this study is relatively limited, focusing on the field of Natural Language Processing. [Conclusions] This paper constructs an automatic innovation sentence recognition engine, preliminarily realizing the practical application of the proposed approach.

[Objective] This study aims to develop an approach for the early identification of patents with high disruptive potential by analyzing patent data in emerging technological domains. [Methods] Based on technology life cycle theory, we construct an indicator system for detecting disruptive technologies and apply it to the quantum computing domain using patent data from the PatSnap database. An ensemble learning model is employed to identify patents with strong disruptive potential. [Results] Leveraging the BERTopic topic modeling framework, we identify five prominent disruptive research fronts: quantum encryption, quantum processors, superconducting qubits, semiconductor-based quantum technologies, and quantum neural networks. These findings demonstrate the effectiveness and feasibility of the proposed method. [Limitations] The empirical analysis is restricted to quantum computing, without extending to other critical technology fields. Moreover, the framework and indicators are solely based on patent data, suggesting opportunities for incorporating additional data sources. [Conclusions] The proposed method provides a systematic approach for the early recognition of highly disruptive patents and the mapping of disruptive research trajectories. The findings offer valuable insights to inform the formulation and implementation of national science and technology strategies.

[Objective] This study proposes a conversational question answering (Q&A) system to address the challenge of natural language queries in bibliographic search systems, which are difficult to accurately map to structured database queries. [Methods] The system employs the Model Context Protocol to achieve seamless integration between the Large Language Model and the external database. To address the issue that exemplar -driven Text-to-SQL generation is susceptible to noise and domain discrepancies, we designed a contrastive learning-based exemplar selection strategy. By fine-tuning the text embedding model to focus more on the syntactic structure and retrieval intent of the query, the quality of similarity ranking is enhanced. We conducted experiments on a constructed bibliographic search semantic parsing dataset to comparatively verify system performance under zero-shot and few-shot conditions. [Results] Compared to the zero-shot setting, the DeepSeek-V3 model achieved an 18.5% increase in SQL execution accuracy in the 5-shot setting, demonstrating the effectiveness of the exemplar selection strategy for the domain-specific Text-to-SQL tasks. [Limitations] Due to the limited coverage of the experimental dataset, the system’s adaptability to cross-domain queries still requires further enhancement. [Conclusions] This research demonstrates the effectiveness of combining large language models with contrastive-learning exemplar selection strategies in intelligent bibliographic search, offering insights for building conversational QA systems in other vertical domains.

[Objective] This paper proposes a federated learning framework embedded with dual-channel attention convolution. It aims to address the challenge of cross-social-network feature extraction to meet privacy protection requirements and to identify social-bot accurately. [Methods] First, we adopted a federated learning framework to integrate datasets across social networks. Second, we embedded dual-channel attention convolution in the local model module to comprehensively extract data features. Third, we used basic convolutional and blockchain technology in the federated aggregation module to integrate parameters from each local model, enabling parameter acquisition and secure storage of optimal model parameters. [Results] Experiments conducted on the TwiBot-20 and Weibo-bot datasets showed that the proposed model achieved an accuracy of 91.63%, precision of 97.10%, recall of 97.14%, and F₁-Score of 96.88%, respectively, demonstrating strong generalization capability. [Limitations] In multi-modal feature extraction, we only considered structured data, text data, and image data; video and audio data were not included. [Conclusions] The proposed model effectively addresses the poor social-bot detection performance caused by insufficient feature extraction and single-source data.

[Objective] This study addresses the limitations of traditional bibliometric tools, which are unsuitable for fine-grained content analysis of literature and incapable of constructing systematic frameworks of research questions and research methods. [Methods] We constructed an ontology to capture the distinctive features of research questions and methodologies in the literature. Then, we identified graph triples by leveraging both local and global contextual information of input sequences. We also utilized the general intelligence of large language models (LLMs) to enhance the accuracy of triple extraction. Finally, we proposed a method for converting natural language into graph query statements, enabling natural language-based knowledge graph querying. [Results] The proposed method improved the F1 score of entity classification and relation extraction by 9.18 and 8.07 percentage points, respectively. The constructed “Problems and Methods” knowledge graph effectively supports popular research and literature association analysis. With the support of LLMs, it enables high-quality graph-based question answering. [Limitations] The ontology design of the knowledge graph remains incomplete and cannot fully support comprehensive literature content analysis. In addition, triplet extraction using the GPT-4o LLM has relatively low efficiency, limiting its applicability for large-scale graph construction. [Conclusions] The proposed knowledge graph can effectively assist scholars in analyzing the current state of research and literature associations from the perspective of problems and methods. Large language models play an important role in the construction and application of academic graphs.

[Objective] This study addresses the challenges of named entity recognition (NER) in ancient Chinese texts, which arise from their condensed linguistic style, substantial variability, and the limited availability of annotated resources. [Methods] We propose a hybrid RAG-LATS framework that integrates an ancient-text knowledge base with AI-Search-enhanced retrieval-augmented generation (RAG). By leveraging the generation, retrieval, reflection, and revision mechanisms of the LATS framework, the approach strengthens the zero-shot NER capabilities of large language models in the domain of ancient texts. [Results] Experiments conducted on the CHisIEC dataset show substantial performance gains: the proposed method improves the Micro F1 score by 14.44 percentage points over the domain-adapted Xunzi-Qwen1.5-7B-Chat model and by 16.99 percentage points over the general-purpose Qwen1.5-7B-Chat model. [Limitations] The design of prompts warrants further refinement, and the computational overhead of LATS may constrain its scalability in large-volume applications. [Conclusions] Retrieval-augmented generation effectively enhances the domain knowledge of large language models, while the LATS framework optimizes the accuracy and coherence of model outputs. Together, these components yield notable improvements in zero-shot NER for ancient Chinese texts.

[Objective] This study introduces classical traditional Chinese medicine (TCM) texts as entity descriptions into representation learning models to improve knowledge graph reasoning performance, addressing the problems of structural sparsity and insufficient utilization of semantic information in TCM knowledge graphs. [Methods] A classical formula knowledge graph was constructed using 398 clauses from Shanghan Lun as the primary corpus, combining manual annotation and automatic extraction in a hybrid process. Relevant passages from Shanghan Lun and Jinkui Yaolue, associated with the graph’s entities, were used as textual descriptions and were encoded using a pre-trained language model. Knowledge graph reasoning experiments were conducted by integrating static BERT representations (BERT-Static) and fine-tuned BERT representations (BERT-Dynamic) with several baseline models, including TransE, DistMult, ComplEx, and SimplE. [Results] On the Jingfang knowledge graph, the reasoning models enhanced with classical text information consistently outperformed their corresponding baseline models in MRR, Hits@1, Hits@3, and Hits@10. In absolute terms, MRR increased from 0.062~0.089 to 0.257~0.270, and Hits@10 increased from 0.148~0.181 to 0.513~0.547. Ablation experiments on input length further indicated that model performance improved as the BERT input length increasing from 16 to 96 Chinese characters, but improvements became marginal beyond this point while training time increased approximately linearly. [Limitations] The experimental data mainly derive from passages in Shanghan Lun and Jinkui Yaolue, and the scale of the knowledge graph remains relatively limited. Moreover, the proposed method only considers the joint representation of entity descriptions and structural triples, while additional structural enhancement mechanisms such as relation paths and rule-based reasoning are not yet incorporated. [Conclusions] Incorporating classical TCM texts as entity descriptions and integrating them with pre-trained language models can effectively enhance knowledge representation and significantly improve reasoning performance in TCM knowledge graphs.