Artificial Intelligence Review最新文献

Legal artificial intelligence (LegalAI) refers to the use of artificial intelligence technologies to automate various legal tasks. Recent advances in large-scale language models have significantly enhanced the capabilities of LegalAI, marking a new stage in its development. In this paper, we present a comprehensive survey of how large language models (LLMs) are reshaping the research paradigm of LegalAI. Beyond improving task performance, LLMs now serve as integral components across the perspectives of data, modeling, and evaluation. We propose a role-based schema that categorizes the involvement of LLMs along these perspectives and use it to systematically review existing studies in three major legal tasks, including legal classification, legal retrieval, and legal generation. Besides, we conduct a detailed quantitative comparison of LLM effectiveness across roles and tasks, and our findings reveal that the impact of LLMs is shaped by both their assigned roles and the nature of the legal tasks.

Visually Rich Documents (VRDs) play a vital role in domains such as academia, finance, healthcare, and marketing, as they convey information through a combination of text, layout, and visual elements. Traditional approaches to extracting information from VRDs rely heavily on expert knowledge and manual annotation, making them labor-intensive and inefficient. Recent advances in deep learning have transformed this landscape by enabling multimodal models that integrate vision, language, and layout features through pretraining, significantly improving information extraction performance. This survey presents a comprehensive overview of deep learning-based frameworks for VRD Content Understanding. We categorize existing methods based on their modeling strategies and downstream tasks, and provide a comparative analysis of key components, including feature representation, fusion techniques, model architectures, and pretraining objectives. Additionally, we highlight the strengths and limitations of each approach and discuss their suitability for different applications. The paper concludes with a discussion of current challenges and emerging trends, offering guidance for future research and practical deployment in real-world scenarios.

This research examines the evolution and integration of large language models within the geospatial domain, exploring both theoretical aspects and practical applications. Using a systematic review guided by the PRISMA methodology, the study investigates GeoAI developments and assesses the impact of foundation models in geospatial contexts. The findings highlight that commercial models demonstrate notable capabilities in interpreting geospatial concepts and generating functional code, although they face limitations concerning accessibility, transparency and reliance on external infrastructures. Smaller, open-source models, adapted through approaches such as fine-tuning and Retrieval-Augmented Generation, are identified as feasible alternatives, providing a balanced solution in terms of accuracy, efficiency and customization. The study emphasizes a need for large-scale, standardized datasets for effective training and evaluation of geospatial models, pointing toward a clear direction for future research. Despite significant advancements, achieving full autonomy of geospatial agents in complex task-solving scenarios remains an unresolved challenge. The future progression of GeoAI will rely heavily on interdisciplinary collaboration and the development of robust, transparent and ethical models capable of supporting real-time decision-making and promoting digital transformation in public administration and related fields.

Sequential recommendation has achieved remarkable progress with self-attentive architectures such as SASRec (Self-Attentive Sequential Recommendation), yet existing approaches often underutilize semantic relationships among items and fail to model temporal dynamics effectively. These limitations reduce the ability of existing systems to capture nuanced user preferences in real-world settings. To address these issues, this work introduces two novel frameworks: one for effective semantic integration and another for a semantic and temporal-aware hybrid embedding generation that enhances the representational capacity of SASRec. These frameworks construct a semantically rich hybrid matrix utilizing Markov State Transition probabilities, Cosine similarity, Personalized PageRank(PPR) based normalization and additionally Time-weighted decay information for the temporal variant. The constructed hybrid matrix is smoothed using a graph convolutional network (GCN) to generate item embeddings, which are then passed to the transformer-based sequential recommendation model SASRec for next-item prediction. Evaluated on three real-world benchmark datasets (MovieLens, Yelp and Amazon Beauty), our proposed temporal variant achieves up to 10.4% improvement in HR@10 and 8.2% in NDCG@10 over SASRec, while maintaining competitive efficiency. Our studies confirm the effectiveness of each component, including semantic graph construction, temporal weighting, and contrastive alignment. These results demonstrate that incorporating semantic and temporal signals into sequential recommenders substantially enhances both accuracy and relevancy of recommendations.

The rapid expansion of complex engineering and real-world optimization problems necessitates the development of efficient, adaptable, and computationally lightweight metaheuristic algorithms. In this study, a novel nature-inspired algorithm called glider snake optimization (GSO) is proposed, which draws behavioral inspiration from the gliding and serpentine locomotion patterns of arboreal snakes to enhance solution exploration and convergence control. The GSO algorithm incorporates a multi-segment movement mechanism, a flexible gliding path generator, and an elite guidance model to ensure effective balance between exploration and exploitation. Extensive experimental validation is conducted using a comprehensive set of 23 classical benchmark functions, high-dimensional test cases (100D, and 500D), the CEC 2019 benchmark suite, and several constrained engineering design problems. The results demonstrate that GSO outperforms or matches 13 state-of-the-art algorithms, including particle swarm optimization (PSO), grey wolf optimizer (GWO), whale optimization algorithm (WOA), and differential evolution (DE) in terms of accuracy, convergence speed, computational cost, and robustness. The algorithm also exhibits exceptional stability across parameter variations, as confirmed through sensitivity analysis and statistical significance testing. These findings highlight the potential of GSO as a powerful and efficient tool for solving complex optimization problems in both theoretical and practical domains. Additionally, GSO achieves leading performance on most benchmark functions, with error reductions of up to 90% compared with competing algorithms. GSO also demonstrates faster convergence and lower variance across repeated trials, confirming its robustness. These quantitative outcomes further reinforce the effectiveness of the proposed algorithm. The MATLAB and Python implementations of GSO are available at https://nimakhodadadi.com.

In recent years, object detection has become a cornerstone of many computer vision applications, relying heavily on the availability of high-quality annotated datasets. However, even widely used benchmarks often suffer from annotation issues such as inaccurate bounding boxes, misclassified objects, and missing labels. These annotation errors, especially localization errors, can greatly affect the training and evaluation of detection models. In this survey, we provide a data-centric and comprehensive review of existing methods for identifying and analyzing errors in object detection datasets. We examine the main components of error detection workflows, including annotation error taxonomies and model-agnostic detection techniques. In addition, we develop a standardized categorization of annotation error types specific to object detection, providing a foundation for consistent analysis and comparison across studies. We also perform manual inspections of selected benchmark datasets to observe and quantify common annotation errors in practice. Moreover, the survey highlights the datasets used for evaluating error detection methods and compares their scope and inherent challenges. Finally, we summarize the types of annotation errors found in existing benchmarks and provide recommendations for future research to enhance dataset quality and reliability in object detection.

Air pollution is a serious global public health threat arising from exposure to toxic ambient pollutants, including particulate matter (PM), sulphur oxides (SOx), nitrogen oxides (NOx), ozone (O₃), carbon monoxide (CO), and ammonia (NH₃). Traditional statistical and deterministic forecasting models often fail to adequately represent nonlinear interactions among multiple pollutants, meteorological drivers, and anthropogenic influences, motivating the growing adoption of deep learning (DL) approaches. This systematic review synthesizes evidence from more than 150 peer-reviewed studies conducted across diverse geographical regions and employing a wide range of DL architectures, including standalone, hybrid, and advanced spatiotemporal models. Using structured quantitative summaries, rank-based performance comparisons, and methodological assessments, the review identifies leading model families, analyzes pollutant- and horizon-specific performance trends, and evaluates robustness and generalizability across spatial and temporal contexts. Overall, DL models generally outperform traditional approaches, particularly when multi-source inputs and spatiotemporal dependencies are explicitly modeled. Nevertheless, the literature remains fragmented, with a strong concentration of studies in data-rich urban regions of Asia, heterogeneous datasets, inconsistent evaluation protocols, limited transparency, and weak external validity. Addressing these limitations requires standardized preprocessing and benchmarking practices, improved explainability and uncertainty quantification, and the development of globally representative datasets. Emerging directions, including hybrid, physics-informed, and generative DL architectures, offer promising pathways to enhance reliability and operational deployment. Collectively, this review provides a comprehensive and critical synthesis of DL-based air quality forecasting, offering actionable insights for researchers, practitioners, and policymakers seeking transparent, generalizable, and policy-relevant prediction systems for environmental management and public health protection.

Online fault diagnosis is crucial for improving the reliability of safety–critical industrial systems. Fault diagnosis becomes increasingly complex if the time-variations, fuzziness, and uncertain causal relationships related to the various internal factors in the system are present. In view of the time-varying working conditions of industrial systems, as well as the fuzzy uncertainty of fault data and knowledge, a C-DUCG (Cubic dynamic uncertain causality graph) approach is proposed for fault diagnosis. Such a solution is intended to help develop a generative cubic causality graph modeling scheme and a FUTURE (fuzzy temporal causality reasoning) algorithm, both of which facilitate the representation and reasoning about complex fault situations (involving temporal causalities and uncertain evidence). In C-DUCG, the strategies of causality simplification and EELA (event-oriented early logical absorption) are proposed to mitigate the complexities of modeling and reasoning. Comparative experiments and a sequence of fault diagnosis tests on a nuclear power plant (NPP) simulator validate the efficiency, recall, accuracy, and interpretability of C-DUCG in large-scale dynamic systems. Experiments reveal that the proposed algorithm achieves 0.62 and 0.999 in terms of recall rate AC@1 and AC@k, respectively, with the average reasoning time being 10 ms, and the average time spent in NPP fault diagnosis tests is 9.42 ms.

Alzheimer’s disease (AD) is a chronic neurological disease and one of the main causes of dementia in around the world. Traditional diagnostic techniques have limits in terms of subjectivity and resource availability, despite the fact that early and accurate identification of AD is essential for efficient supervision. Deep learning (DL) and Machine Learning (ML) have emerged as powerful tools in medical imaging and have shown promising results in AD detection. This review provides a comprehensive analysis of the latest developments in ML and DL for AD diagnosis, covering essential data sets such as Alzheimer’s Disease Neuroimaging Initiative (ADNI), Open Access Series of Imaging Studies (OASIS), and Australian Imaging, Biomarkers and Lifestyle Study of Ageing (AIBL), as well as preprocessing techniques that enhance data quality. Here, we review some of the significant AD studies and investigate how ML and DL might assist researchers in making an early diagnosis more accurate.

Diabetes mellitus is one of the most pressing global health challenges, and early prediction is critical to reducing complications, mortality, and healthcare costs. Conventional diagnostic tools remain limited, as they often rely on a small set of biomarkers and fail to capture lifestyle, genetic, or environmental risk factors. This review systematically evaluates how artificial intelligence (AI), including machine learning (ML) and deep learning (DL), enhances diabetes prediction by integrating multimodal data and improving clinical interpretability. Following PRISMA 2020 guidelines, a systematic search was conducted across PubMed, Scopus, IEEE Xplore, Web of Science, and ACM Digital Library for studies published between 2010 and 2024. Inclusion criteria required AI-based diabetes prediction models with reported performance metrics. From 2134 records, 155 studies met the criteria and were synthesized. AI models consistently outperformed conventional approaches (60–75% accuracy), Ensemble methods such as Random Forests and XGBoost achieved accuracy of 85–90% and AUC-ROC values > 0.90 (Abnoosian in BMC Bioinf 24:373, 2023. https://doi.org/10.1186/s12859-023-05465-z; Chen and Guestrin in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, Association for Computing Machinery, 2016. https://doi.org/10.1145/2939672.2939785). DL architectures showed notable strengths in unstructured data: CNNs achieved AUC > 0.95 in retinal image analysis (Gulshan et al. in JAMA 316(22):2402–2410, 2016. https://doi.org/10.1001/jama.2016.17216; Kermany et al. in Cell 172(5):1122–1131, 2018. https://doi.org/10.1016/j.cell.2018.02.010), while LSTMs reached 85–90% accuracy in continuous glucose monitoring predictions (De Bois et al. in Prediction-coherent LSTM-based recurrent neural network for safer glucose predictions in diabetic people, arXiv:2009.03722, 2020; Bian et al. in PLoS ONE 19(9):e0310084, 2024. https://doi.org/10.1371/journal.pone.0310084). Emerging methods such as federated learning enable privacy-preserving cross-institutional collaboration with comparable accuracy (~ 88%), while explainable AI techniques (e.g., SHAP, LIME, attention mechanisms) enhance transparency and clinical trust. Real-world case studies demonstrate improvements in early detection, reduced hospitalizations, and increased patient engagement when AI models are integrated into EHRs or mobile health apps. This review contributes a novel synthesis by combining methodological insights with clinical applications. The central takeaway is that AI-driven diabetes prediction offers significant advantages over traditional methods, but challenges in data quality, generalizability, fairness, and regulatory compliance remain. Addressing these will be essential to ensure safe, equitable, and clinically meaningful adoption in healthcare practice.