The onset of ChatGPT, Sora, Claude-3, and the like, has brought about the era of AIGC for text, images, and videos. Literature review, scientometric analysis, and S&T trends analysis have also being rapidly taken over by AI tools, thus making traditional knowledge services (KS) falling into a “low-quality productivity trap”. It is difficult to develop new quality productive forces with competitive vitality and resilience by only using AI to optimize the execution efficiency of traditional KS business logic. AI, as illustrated by Large Language Models (LLM), has broken the reductionist “research model” that disassembles complex phenomena and systems into individual parts to study and solve, and the turing computing model that pursues deterministic computing, therefore able to handle the dimensional disaster from the combinatorial explosion of complex multi-interactive systems. This helps us to truly take “complex problems, dynamic decision conditions, and selective operational solutions” as the goal of KS, and provides users with decision intelligence. This may be the starting point in the search for new quality productive forces in KS. But it is imperative to ask “what is the real problem” from the point of First Principle. Starting from the fundamental needs of decision users of KS, we need to think clearly about what KS should do, can do, and must do. Admittedly, when problems of decision-makers ask for various data or information analysis, what they really need is to answer is not just “what is” but “why is so and what can/should I do” in their S&T planning, organizing, resourcing, evaluating, etc., under their specific conditions. If so, KS should now be positioned as a “user decision-making productivity service”, focusing on Policy for S&T (P4ST), hence transforming KS from the literature-oriented or data-oriented or indicator-oriented to user-problem/solution-oriented models, and from data or computational intelligence to cognitive and decision intelligence. Based on several examples, this paper proposes a generalized decision-making genomic model for AI-empowered Policy for Science & Technology (AI4P4ST). The model consists of an Agent axis (multi-levels from individuals to nations), an Action axis (planning, organizing, budgeting, evaluation, etc.), and an Outcome axis (plans, institutions, teams, projects, papers, patents, products, etc.). Use of this model supports intelligent decision-making analysis under the dynamics of complex systems. With multiple combinations of variables that interact in known or unknown ways, we can perform multi-modal cross-scale modeling and analysis of multi-dimensional multi-variates, continuously adjusting to approximate possible solutions with quantifiable uncertainties, so that decision-makers can select for a decision. AI4P4ST analysis can progressively implement the P4ST analysis pipelines that supports the dynamics of complex systems. The LLM Prompt Engineering and its many augmented models can be used to build an AI4P4ST Chain of Analyses. In addition, technologies such as AI Agents, Multi-Agents Models, and Mixture of Experts (MoE) models, as well as mechanisms such as LangChain or GPTSwarm, can be employed to support AI-enabled application processes that combine multiple LLMs and specialized tools, thus enabling intelligent processes such as planning, prediction, experimentation, verification, and analysis for AI4P4ST. Of course, AI4P4ST still faces challenges from complex data environments and complex social dynamics, including multi-modal heterogeneous data environments, boundary uncertainty, strong game adversariality, difficulties in handling critical states, and weak counterfactual reasoning. This may require a combination of knowledge-based intelligence modeling, simulation and prediction based on the complex system dynamics, and data-based LLM modeling, and the use of LLM models to plan, coordinate, and support these modeling and analysis.
[Objective] This paper summarizes the knowledge graph completion methods based on graph neural network through research and literature review. [Coverage] With “knowledge graph completion” as search terms to retrieve literature from the Web of Science, DBLP and CNKI, a total of 79 representative literature were screened out for review. [Methods] Based on the model structure, three knowledge graph completion methods based on graph neural networks were summarized, including graph convolutional neural networks, graph attention networks, and graph auto encoder. [Results] Using common data sets and evaluation indicators for knowledge graph completion tasks, the effects of various models were comparatively analyzed in terms of MRR, MR, Hit@k and other performance evaluations, and prospects for future research were suggested. [Limitations] In the comparison of experimental results, only the evaluation results of some widely used models on the FB15K-237 and WN18RR datasets are discussed, the comparison of all models on the same dataset is lacking. [Conclusions] Compared with the representation learning model and the neural network model, the graph neural network model has better performance, but it still faces difficulties such as high complexity of model relationships, over-smoothness, and poor scalability and universality.
[Objective] This paper reviews the research progress on semantic novelty in China and abroad. It explores relevant techniques and provides references for future studies. [Coverage] We used keywords such as “novelty of the literature”, “semantic novelty”, “literature novelty”, and search expressions like “semantic novelty and literature evaluation” to retrieve literature from Web of Science, Elsevier, Springer, Google Scholar, as well as Chinese databases like CNKI, Wanfang, and VIP. A total of 70 representative literature were selected for review. [Methods] We summarized research on semantic novelty, focusing on the definition of novelty, evaluation indicators, and different evaluation methods. We also discussed the current development status and future trends of evaluating semantic novelty in scientific literature. [Results] Semantic novelty evaluation has gradually received widespread attention from the academic community. Related studies have evaluated semantic content without establishing a unified measurement index. [Limitations] Existing evaluation of literature novelty mainly focuses on external features. Fewer research papers directly addressed semantic novelty, limiting support for reviews. [Conclusions] The evaluation of semantic novelty in scientific literature fundamentally lies in the novelty of semantic content. Quantitative research has become the mainstream method, but the calculation method of evaluation indicators needs to be clarified. Future studies on novelty evaluation should combine qualitative and quantitative methods for more comprehensive evaluations.
[Objective] This study constructs a predictive model for depression severity and explores its interpretability issues. We aim to improve the automated depression detection model’s reliability and practicality by analyzing Internet user-generated content. [Methods] First, we built a corpus by collecting depression-related medical consultations from the Good Doctor Online platform. Then, we extracted patients’ psychological features using C-LIWC, a psychology lexicon. Third, we predicted the patients’ conditions with the Gradient Boosting Tree algorithm. The study also incorporated the explainable machine learning method SHAP to interpret the new model. Through SHAP’s unique visualizations, we analyzed the complex relationship between patients’ age, gender, cognition, emotions, perceptions, social / family contexts, personal gains or losses, and the occurrence of depression. [Results] The psychological state of depression patients provided feedback on their condition. Utilizing psychological features extracted from consultation records effectively detected severe depression, with an accuracy of 86%. The SHAP reveals multiple effects of patients’ psychological features on depression. [Limitations] Limited by the corpus, predictions of depression severity were based only on single consultation records. Additionally, the model features were based on psychological dictionaries, while more elements related to the risk of depression could be included in the future. [Conclusions] Factors influencing the occurrence and development of depression are complex. Individual differences result in different effects of various characteristics on disease prediction. Building an automated diagnostic model for depression should focus on the model’s accuracy and enhance understanding of the model’s predictions.
[Objective] This study conducts sentiment analysis using the emotional knowledge contained in the syntactic structures of texts from online health communities. We propose an online health community sentiment analysis model, WoBEK-GAT, based on emotional enhancement and knowledge fusion. [Methods] Firstly, we utilized WoBERT Plus for dynamic word embedding. Then, we extracted semantic features using CNN and BiLSTM. Finally, we fully integrated key syntactic information from pruned dependency trees with external emotional knowledge through sentiment enhancement and knowledge fusion strategies. We fed these inputs into the GAT to output sentiment categories. [Results] We conducted comparative experiments on a constructed Chinese dataset. The proposed model’s MacroF1 value reached 88.48%. It was 15.49%, 14.15%, and 13.15% over baseline models CNN, BiLSTM, and GAT, respectively. [Limitations] We should have considered sentiment knowledge in multimodal information such as pictures and speeches. [Conclusions] The proposed model could effectively improve sentiment analysis capability.
[Objective] This study explores the mechanism of threat assessment on users’ willingness to engage in defensive privacy protection behaviors. It helps companies make reasonable privacy management decisions and foster a healthy corporate digital ecosystem. [Methods] Based on the protection motivation theory and focusing on threat assessment, we introduced “information privacy anxiety” as an emotional mediating variable. Then, we used the information sensitivity of the context as a moderating variable to construct a model for the impact mechanism of threat assessment on users’ defensive intentions. We used the SEM-PLS to empirically analyze 183 financial and 200 e-commerce context datasets. [Results] Information privacy anxiety is a critical emotional factor influencing users’ willingness to defensive privacy protection behaviors. It plays a partial mediating role between perceived threats and defensive intentions. The information sensitivity of the context positively moderates the relationship between information privacy anxiety and defensive willingness. The information sensitivity of the context only has a moderating effect on the relationship between perceived vulnerability and threat. In contrast, it did not moderate the relationship between perceived severity and threat. [Limitations] This study explores willingness rather than actual behaviors. Regarding information sensitivity comparison, we only chose representative finance and e-commerce contexts. [Conclusions] This study advances the protection motivation theory, providing theoretical guidance for business to adopt appropriate management measures to reduce users’ defensive privacy protection behaviors.
[Objective] This paper aims to improve the low accuracy of sentiment classification using the pre-trained model with insufficient samples.[Methods] We proposed a prompt learning enhanced sentiment classification algorithm Pe(prompt ensemble)-RoBERTa. It modified the RoBERTa model with integrated prompts different from the traditional fine-tuning methods. The new model could understand the downstream tasks and extract the text’s sentiment features. [Results] We examined the model on several publicly accessible Chinese and English datasets. The average sentiment classification accuracy of the model reached 93.2% with fewer samples. Compared with fine-tuned and discrete prompts, our new model’s accuracy improved by 13.8% and 8.1%, respectively. [Limitations] The proposed model only processes texts for the sentiment dichotomization tasks. It did not involve the more fine-grained sentiment classification tasks. [Conclusions] The Pe-RoBERTa model can extract text sentiment features and achieve high accuracy in sentiment classification tasks.
[Objective] This paper constructs a multimodal topic model combining text and images in news based on multimodal learning methods. It aims to uncover latent topics in the news automatically. [Methods] We constructed a model incorporating word embedding for topics from texts and images. It uses multimodal joint representation learning and coordinate representation learning for feature fusion. We conducted visual analysis for the discovered multimodal news topics. Finally, we examined the new model on the N15News dataset. [Results] Compared to Label-ETM using only text features, the multimodal topic modeling approach can achieve better topic interpretability and diversity. This suggests that the multimodal topic modeling approach is feasible. [Limitations] We assume images and text in news are semantically and thematically related. Multimodal fusion methods need to be improved in weakly related and irrelevant domains. [Conclusions] Multimodal topic modeling can discover connections between different modal data and improve the diversity of discovered topics.
[Objective] This paper proposes a new deep clustering model (ISCCL) for texts based on SCCL, aiming to improve its performance in clustering tasks. [Methods] First, the ISCCL model utilized sentence vector pre-trained models to perform data augmentation and encoding to obtain two sets of augmented representations of the input texts. Then, we added two layers of nonlinear networks to the SCCL model. It reduced the augmented representations to a cluster feature space with dimensions equal to the number of clusters. Third, we constructed positive and negative cluster pairs from the perspective of column space for contrastive learning. It guided the model to explore valuable features for clustering tasks and reduce the impact of false positive samples. [Results] In five benchmark datasets, including AgNews, Biomedical, StackOverflow, 20NewsGroups, and zh10, the clustering accuracy of the ISCCL model reached 88.89%, 48.74%, 78.17%, 56.97%, and 86.42%, respectively, which is an improvement of 0.69% to 2.67% compared to the SCCL model. [Limitations] The dimension of the cluster feature space needs to be pre-set (the same as the clustering number K value). However, it is often difficult to determine the specific cluster number of the original data, and adjustments should be made according to the dataset. [Conclusions] The ISCCL model can effectively extract cluster features and improve the deep clustering performance on texts.
[Objective] This paper proposes a new Chinese abstract generation method integrating content and image features. It aims to improve the performance of existing methods based on text features. [Methods] First, we used the BERT to extract text features and used ResNet to extract image features. Then, we utilized these features to complement and validate each other. Third, we fused the two modal features with the attention mechanism. Finally, we inputted the fused features into a pointer generation network to generate higher-quality Chinese abstracts. [Results] Compared to models solely relying on single text modality, the proposed method showed improvements of 1.9%, 1.3%, and 1.4% on ROUGE-1, ROUGE-2, and ROUGE-L metrics, respectively. [Limitations] The experimental data were primarily retrieved from the news domain, and the model’s effectiveness in other fields remains to be verified. [Conclusions] Incorporating image information allows the fused features to retain more important information. It helps the model identify the key content better and makes the generated abstracts more comprehensive and readable.
[Objective] This paper designs a retrieval model to explore the interpretability of question-answering tasks. It examines the reasoning processes of these reading comprehension models and improves sentence relevance in traditional unsupervised retrieval algorithms. [Methods] We proposed a new unsupervised retrieval model ISR, which integrated modules of Pearson correlation coefficient, GloVe word embeddings, and IDF weighting. The ISR model conducted fine-grained retrieval of inference sentences through multi-round iterations. [Results] The proposed model’s P, R, and F1 metrics were 2.4%, 1.8%, and 2.1% higher than the MSSwQ model on the MultiRC dataset. Its P, R, and F1 metrics were 4.8%, 2.6%, and 3.7% higher than the MSSwQ on the HotPotQA dataset. [Limitations] There might be excessive matching issues due to the model’s retrieval matching mechanism. [Conclusions] The proposed model improves the accuracy of retrieval inference sentences, which can be effectively applied to the question-answering tasks.
[Objective] This study proposes a deep learning-based recommendation model for research teams to meet recruitment needs and improve recommendation efficiency. [Methods] Firstly, we applied the self-attention mechanism to learn the semantic representation of teams. Then, we employed the neural collaborative filtering model to study the nonlinear relationship between teams and researchers. Finally, we obtained the degree of fit between teams and individuals as the basis for recommendation. [Results] Compared with the baseline models, the proposed one increased the recommendation accuracy and F1 value by 10.22% and 10.25%, respectively, on public datasets. It performed exceptionally well in real-world recommendation scenarios. [Limitations] The parameter size of the deep learning model is relatively small, leaving room for optimization. [Conclusions] The proposed model can effectively enhance the efficiency of recruiting researchers, helping research service institutions improve their services and satisfy the needs of research teams.
[Objective] To inherit and promote seal culture and enhance the recognition of seals in complex scenarios, this study structurally displays the recognition results and related knowledge using knowledge graphs and visualization techniques. [Methods] We proposed a deep learning model integrating multiple features. First, we extracted the seal images’ color, edge, and grayscale feature maps. Then, we input these feature maps into the deep learning model for recognition. Finally, we compared the recognition results with the nodes in the knowledge graph and visualized the related knowledge. [Results] The study collected and annotated seals from 13 calligraphy and painting works, including “The Cold Food Observance”, with two selected works as the test set. Compared with the VGG16 model, our new model’s precision (P), recall (R), and F1 score improved by 28.40%, 28.67%, and 28.54%, respectively. Without integrating multiple features, the P, R, and F1 values decreased by 24.30%, 20.16%, and 22.74%, respectively. [Limitations] The proposed model can only extract and recognize global features of seals, lacking the ability to identify and infer their local semantic information. [Conclusions] The proposed method has a good effect on seal recognition tasks, where multi-dimensional feature maps can enhance the model’s recognition ability and robustness in complex cases.
[Objective] To fill the current deficiency in visual emotional analysis research, a ResNet34-based improved emotion analysis model was constructed to analyze and improve the accuracy of image emotion classification. [Methods] Firstly, a visual emotion recognition model was established based on the ResNet34. Then, by integrating the CBAM module and Non-Local module, emotion features were learned and represented. Finally, the above model was used to classify and recognize emotional features, and compared with VGG16 and ResNet50 models. [Results] The recognition effect of the constructed model was verified through experiments, and the research results showed that the accuracy, precision, recall, and F1 score of the model reached 84.42%, 84.10%, 83.70%, and 83.80% respectively. Compared with the baseline models of the VGG16 and ResNet50, the accuracy of the proposed model was improved by 4.17% and 3.44%, and the F1 score was improved by 4.20% and 3.30%. [Limitations] The scale of the test dataset is relatively small, the effectiveness of annotation was not calculated using metrics such as the Pearson correlation coefficient, and a comparison was not made with visual-based emotion classification algorithms. [Conclusions] From the perspective of visual emotional analysis, optimizing the emotion recognition model supplements the analysis mode of emotional computation, providing support for the extraction and analysis of emotional features in public opinion information.
