Purpose Retrieval-augmented generation (RAG) systems integrate information retrieval with generative language models to improve the relevance, accuracy and explainability of AI-driven responses. This study evaluates how different configurations of embedding and generative models nfluence the performance of RAGpipelines for knowledge management (KM) scenarios. Design/methodology/approach The study combines a broad benchmark of embedding and generation components with a contemporary open-weight comparison centered on Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.3 and Gemma-2-9B-It. Retrieval configurations are evaluated through recall, latency and storage trade-offs, while generation quality is assessed using ROUGE-L, exact match (EM), token-level F1, BERTScore F1, semantic similarity, answer relevance and faithfulness. The benchmark also includes complementary evaluation on SQuAD and HotpotQA, grounded prompting, abstention prompting, error analysis and long-context stress testing. Findings Retrieval quality remained the main determinant of end-to-end RAG quality. The strongest shared retrieval setup combined all-mpnet-base-v2, 256-token chunking with 64-token overlap and top-1 retrieval, reaching Recall@1 = 0.938. Among the open-weight generators, Gemma-2-9B-It achieved the strongest lexical and semantic matching, with its best grounded-abstain configuration reaching ROUGE-L = 0.631, EM = 0.456, token-F1 = 0.631 and BERTScore F1 = 0.767. Llama-3-8B-Instruct produced the strongest faithfulness score in the best grounded setting (0.241), while Mistral-7B-Instruct-v0.3 occupied a more conservative operating point with lower answer matching but stronger abstention behavior. HNSW matched exact-search quality for equivalent retrieval configurations while reducing query latency. Practical implications The findings support retrieval chunking, top-1 retrieval and grounded prompting as robust design choices for question-answering-oriented RAG. They also suggest that safer abstention-oriented prompting should be treated as a different operating point rather than as a universal default. Social implications More reliable RAG systems can improve access to institutional knowledge, support organizational learning and reduce barriers to expertise discovery, especially when system designs balance quality, latency and computational cost. Originality/value The paper contributes a component-level benchmark for RAG in KM settings, richer evaluation dimensions and a more explicit treatment of retrieval/generation trade-offs across historical and contemporary open-weight baselines. The design narrows practical claims to what is supported by multi-dataset evidence, error analysis and long-context testing.
Purpose This work introduces a novel model for understanding the nuanced dynamics of the brain signal during transition periods between stimuli presented to patients in functional magnetic resonance imaging (fMRI). Design/methodology/approach This new mathematical method takes as input the voxel data from the relevant brain region, applies dimension reduction and fits a spline through each feature. The features from the constructed spline, such as the mean first derivative, mean absolute second derivative, etc., are used to classify the state per sliding window by labeling it as either transition or non-transition between stimuli. For the classification task, a random forest classifier is used in an experiment with data from six participants, and with 5-fold time series cross-validation. Findings The results from data of 6 participants from the Haxby dataset show an impressive 68% accuracy on average for the random forest classifier with splines, compared to the 60% baseline random forest classifier trained using a wavelet transform and wave reconstruction approach. Originality/value This work introduces a new mathematical framework for detecting cognitive state transitions in fMRI based on geometric features, demonstrating a substantial improvement over baseline feature representations.
Purpose Post harvest losses (PHLs) pose significant challenges to food security and farmers' livelihoods. Within the complex supply chain, truck transportation emerges as a primary contributor to these losses. This study introduces a decoupled methodology, combining process mining and data mining, to predict and mitigate PHL. It is achieved by combining insights from process mining and advanced data mining techniques. By leveraging transportation logistics, environmental conditions and other relevant environmental and operational factors, this work aims to provide actionable insights for reducing waste and enhancing the efficiency of the post-harvest supply chain. Design/methodology/approach This study introduces a novel framework for PHL prediction utilizing decoupled methodologies that integrate process mining and data mining. We leverage transportation data from our self-developed Raw Material Tracking (RMT) mobile application and environmental information. Initially, process mining refines noisy transportation data from our Raw Material Tracking (RMT) mobile application to extract process-aware features. Subsequently, these refined features feed into a stack ensemble classification model for PHL prediction. We utilize self-supervised clustering to determine optimal high/low loss thresholds. Following the prediction, association rule mining identifies key patterns linked to these high/low loss outcomes. This modular design that integrates process insights with predictive and pattern analysis underpins our decoupled approach. Findings The analysis result yielded several critical insights into PHL. First, important feature analysis clearly identified the most influential attributes for PHL prediction. We validated our clustering results by experimenting with various loss thresholds to determine the optimal cutoff point for distinguishing high from low losses. In addition, through associative pattern analysis, we uncover that environmental factors such as temperature and humidity, combined with transportation-related data, are key determinants of PHL. These findings offer actionable insights into the precise conditions and process elements contributing to PHL. Research limitations/implications The current analysis is geographically constrained to a single city, that is Surabaya. It is necessary to provide broader validation across diverse settings to enhance generalizability. Furthermore, reliance on city-level local weather data limits micro-level accuracy along truck routes; thus, future work should integrate more granular, route-specific meteorological data. Additionally, incorporating more relevant features, such as product characteristics and other operational aspects, is essential to understand their influence on PHL and further refine prediction accuracy. Practical implications The advantage of our decoupled methodology lies in its ability to isolate and optimize distinct analytical steps from process data refinement to predictive modeling and pattern discovery. This modularity allows for robust insights that would be challenging to achieve with monolithic approaches. By utilizing this framework, stakeholders can make more informed decisions regarding storage conditions and delivery schedules that lead to a significant reduction in PHL, improved supply chain efficiency and enhanced profitability. Originality/value This study advances PHL prediction by introducing a novel data-driven framework. Unlike prior research works that predominantly rely on static tabular datasets, this study incorporates both sequential data and tabular data within a robust machine learning pipeline. This dual data integration, particularly leveraging process-aware sequential insights, enhances predictive accuracy and offers a better understanding of loss factors. Furthermore, our methodology is highlighted by its ability to equip practitioners with reliable, data-derived loss thresholds to enable more precise loss classification and targeted mitigation strategies.
PurposeAlthough citation-based indicators are sometimes used to help evaluate the quality of papers in conference-based fields, conferences seem to be less systematically indexed in the major citation indexes, and the value of citation counts as research quality indicators for them is unknown. In response, this article investigates whether ChatGPT might provide a suitable alternative.Design/methodology/approachChatGPT was used to assign a quality score to conference papers from 2020 from the 23 narrow fields that are partly conference-based in the sense of having at least 30% conference papers indexed in Scopus. The results were compared with citation counts and conference rankings.FindingsChatGPT research quality score predictions based on paper titles and abstracts alone correlated positively and statistically significantly with citation rates in all fields at the paper level and mostly at the conference level. Expert-based citation-informed conference rankings conformed slightly more closely with geometric mean citation rates than with mean ChatGPT scores.Research limitations/implicationsNo direct measure of paper research quality was used.Originality/valueWhilst the evidence tends to support the value of both citations and ChatGPT as research quality indicators, it suggests that citations may be better, at least for older research, and that ChatGPT is a reasonable alternative for research that is too new to have attracted many citations, at least in conference-based fields.
Purpose This paper’s purpose is to enhance multivariate time series forecasting accuracy by introducing a novel preprocessing step: leveraging advanced image deep learning denoising models via a transfer learning approach. We specifically evaluate denoising convolutional neural network (DnCNN), deep residual neural network (ResDNN) and deep residual network (DRNet) on “individual household electric power consumption” and “the Rainforest Automation Energy” (RAE) datasets. The study aims to demonstrate that denoising significantly improves predictive accuracy and computational efficiency. Our findings highlight ResDNN as the superior denoiser, reducing RMSE by 20–30% and training time by up to 40%. This research underscores the potential of advanced denoising to refine real-world time series data for more accurate and efficient forecasting. Design/methodology/approach Our methodology integrates advanced image deep learning denoising models as a crucial preprocessing step for multivariate time series forecasting. We apply three state-of-the-art denoisers – DnCNN, ResDNN and DRNet – to raw time series data from “individual household electric power consumption” and “the RAE” datasets. The denoised time series are then fed into a subsequent forecasting model (e.g. long short-term memory (LSTM) and gated recurrent unit, etc.; though not explicitly named in the abstract, it is implied by “forecasting model”) to evaluate the enhancement. Through comparative analysis using root mean square error (RMSE) and training time, we assess each denoiser’s efficacy in improving predictive accuracy and computational efficiency compared to a baseline without denoising. This transfer learning approach aims to refine input data quality, leading to superior forecasting performance. Findings Our findings demonstrate the significant benefits of integrating advanced deep learning denoising as a preprocessing step for multivariate time series forecasting. Among the evaluated models, ResDNN consistently emerged as the top performer across “individual household electric power consumption” and “the RAE” datasets. Specifically, ResDNN achieved a notable reduction in RMSE values, ranging from 20% to 30% compared to forecasts made without any denoising. Furthermore, ResDNN also exhibited substantial improvements in computational efficiency, leading to reductions in training time of up to 40%. These results underscore the promise of such methodologies for enhancing real-world time series forecasting. Originality/value This study introduces a novel approach to multivariate time series forecasting by pioneering the use of pretrained image deep learning denoisers as a preprocessing step. To our knowledge, this is the first time that models like DnCNN, ResDNN and DRNet, originally developed for image denoising, have been systematically applied and evaluated in this specific context. We leverage their inherent ability to extract clean signals from noisy data, traditionally in visual domains and transfer this capability to time series. This unique methodological integration offers a fresh perspective on enhancing data quality for forecasting models, moving beyond conventional statistical denoising methods and opening new avenues for cross-domain application of deep learning advancements.
PurposeEnhancing Question-Answering accuracy on new law documents by using Retrieval-Augmented Generation (RAG)-Large Language Models . Design/methodology/approachThe proposed method is built on a RAG design for Indonesian new law documents, incorporating a specific data chunking method and a reranker model. FindingsThe proposed method, which segments legal data by focusing on the document title, article, and paragraph, outperforms the sequential chunking method in terms of accuracy. Originality/valueThis research demonstrates that the completeness of a sentence in the data set used during the retrieval phase of RAG significantly affects the accuracy of RAG responses. This is particularly true for legal documents, which are closely tied to the titles of articles and the wording of each legal provision. The relationship between the article titles and the content enhances the context-awareness of each data chunking process.
PurposeWith the increasing popularity of mobile devices in recent years, users can share their check-in behavior on social networks at any time or any place. However, with the increasing number of users sharing information on social networks and the Internet, users now find it challenging to find the information they need. Many studies have proposed Point-of-Interest (POI) recommender systems to address information overload, aiming to predict POIs based on POI tags, categories, geographic location, and users' check-in times. However, these features are too specific to represent the characteristics of users and POIs.Design/methodology/approachThis work proposes a POI recommendation method based on user reviews and geographic area features (PRRG) to predict which POIs users may be interested in. The research framework includes review analysis and POI area analysis. The review analysis extracts topics, sentiment, and semantic features from user reviews to represent user preferences and POI features. The POI area analysis divides POIs into distinct areas and calculates area weights based on users' movement patterns. Finally, the weighted matrix factorization method is used to predict the POI ratings. It captures the semantics of user preferences expressed in text and users' sequential mobility traces.FindingsThe proposed method can extract various features to represent user preferences and POI features and analyze the importance of the POI area to users based on their movement patterns, thereby enhancing recommendation accuracy. The experimental results demonstrate that the proposed method outperforms other methods and significantly enhances the performance of recommendations.Originality/valueThe proposed POI recommendation integrates review analysis with geographic area analysis. Multiple features are extracted from user reviews, movement tracks, and geographic areas to represent user preferences and POI features. Feature extraction can effectively improve the accuracy of POI recommendations.
PurposeThe purpose of this study is to develop a hybrid approach for document image quality assessment (DIQA) that effectively integrates traditional handcrafted features with deep learning-based features to improve accuracy, robustness and generalization across various document types and degradations.Design/methodology/approachThe proposed method extracts handcrafted features such as Local Binary Patterns, Gray-Level Co-occurrence Matrix, Gabor filters, Scale-Invariant Feature Transform and Haar-like features, as well as deep features using pre-trained convolutional neural networks (EfficientNet, Visual Geometry Group Network (VGGNet), InceptionV3 and ResNet). These features are fused into a composite vector, followed by dimensionality reduction using Principal Component Analysis. A Random Forest regression model is then trained to predict image quality scores. The method is evaluated on the Smartdoc-QA dataset.FindingsExperimental results demonstrate that the proposed method achieves higher correlation with ground-truth scores than several state-of-the-art methods. The integration of traditional and deep features enhances robustness against various distortions such as blur, noise, skew and lighting variations, and results in improved Pearson Linear Correlation Coefficient and Spearman Rank Order Correlation Coefficient metrics.Originality/valueThis paper presents a novel hybrid DIQA framework that leverages both low-level structural/textural cues and high-level semantic representations. It is among the first to systematically combine shallow and deep features with dimensionality reduction and regression to form an end-to-end DIQA system. The approach is scalable, accurate and applicable to real-world document processing scenarios.
PurposeWe are the first to develop the investor sentiment indicator in order to evaluate the level of volatility in the Vietnamese stock market. We then identify connections, specifically differentiating between lagged linkages and contemporaneous connections between investor sentiment and the stock marketplace.Design/methodology/approachOur analysis uses the R2-decomposed linkage methodology to identify connections, specifically differentiating between lagged and contemporaneous linkages. The return transmission mechanism between the Investor Sentiment Index (SI), FPT Corporation (FPT), Hoa Phat Group JSC (HPG), Lam Dong Pharmaceutical JSC (LDP), Gemadept Corp (GMD), Ducgiang Chemicals & Detergent Powder Joint Co (DGC) and Danang Airport Services JSC (MAS) is examined using the above novel method. Our analysis covered January 1, 2017, to February 1, 2024.FindingsLagged linkage exhibits a significantly more pronounced influence, particularly on investor sentiment. The role of the net shock receiver is prominently evident during the final three-quarters of 2019, the final three-quarters of 2020 and towards the conclusion of the observed period. Conversely, the opposite role is observed during the remaining periods. The stock market reached its peak during the COVID-19 epidemic, fostering optimistic confidence among investors. However, during the recessionary phase of 2022-2023, while blue-chip stocks remained relatively unaffected by investors' pessimism, mid-cap and high-risk groups exerted an adverse effect on investor sentiment. The empirical analysis also reveals temporal heterogeneity in the net spillover roles of the examined equities, characterised by distinct regime shifts where entities such as DGC and MAS transitioned from shock receivers to transmitters, while FPT and HPG displayed complex, model-dependent connectivity patterns. Furthermore, the interaction with investor sentiment exhibits a structural evolution, wherein major equities largely drove sentiment during the COVID-19 expansion (2020-2021) but became susceptible to market mood during the 2022-2023 downturn, highlighting a divergence where blue-chip assets remained insulated from negative sentiment relative to mid-cap and high-risk segments.Originality/valueWe investigate a unique chance to explore investor confidence as well as stock market performance in extreme conditions. This paper proposed an R2 decomposed linkage method to evaluate the instability interdependence of several kinds of markets. Our article also marks the pioneer of exploring the influences of unforeseeable occurrences, namely the COVID-19 pandemic and the Russia-Ukraine war, on the interconnection of instability in the stock market's volatility and investor emotion.
PurposeThis paper provides an extensive overview of existing datasets designed for the AI-based detection of false information, and it presents two new datasets. These datasets contain disinformation cases on climate change and the Russian invasion of Ukraine.Design/methodology/approachThe Debunking datasets were built using a systematic and methodologically rigorous approach to ensure data accuracy and representativeness. Data collection involved the amalgamation of multiple sources and put together a wide range of disinformation and real claims and related multimedia content.FindingsEarly tests demonstrate the potential of these datasets to improve the accuracy of fake news classification in these domains. The study concludes with a discussion on future developments and the ethical implications of AI-based fact-checking systems.Originality/valueThe study produced high-quality, domain-specific datasets that address the deficiencies of existing resources. First, a fundamental aspect of the Debunking datasets is domain-specific curation. In fact, unlike general fake news datasets, the AI4Debunk project focuses on two critical areas where disinformation has a profound societal impact. Additionally, Debunking datasets consist of multimodal data. That means the datasets contain images, videos and social media metadata in addition to text statements. Another major strength of the Debunking datasets is their intensive annotation process. The data labeling process is conducted by experienced fact-checkers so that the disinformation claims are accurately labelled.
PurposeObject detection and instance segmentation play an important role in autonomous driving, where vehicles must perceive their surroundings reliably. In practice, these tasks are commonly addressed using separate models, which increases both training complexity and deployment cost. To overcome this issue, we propose UniPercepNet-S, a lightweight dual-task framework inspired by YOLOF that brings detection and segmentation into a single unified network, aiming to support real-time perception in resource-constrained environments.Design/methodology/approachUniPercepNet-S follows a YOLOF-style one-level detection design and strengthens the backbone with a channel attention module to improve feature quality. To enable instance segmentation, we add a simple yet efficient mask prediction branch that operates directly on detected objects while keeping computation low. We evaluate the proposed framework on MS COCO and BDD100 K, covering both general object segmentation and autonomous-driving-oriented scenarios.FindingsThe proposed UniPercepNet-S achieves a mask AP of 38.0 on MS COCO, placing it among the top-performing entries in the COCO Detection Challenge for segmentation tasks. On BDD100 K, which reflects real-world driving conditions, the model reaches an AP of 20.3, showing that it generalizes well across different datasets. These results suggest that UniPercepNet-S can deliver accurate detection and segmentation while remaining suitable for real-time use.Originality/valueThis work contributes a unified and lightweight one-level framework that performs object detection and instance segmentation simultaneously, avoiding the need for heavy multi-scale architectures or separate task-specific models. By combining attention-enhanced representations with an efficient segmentation branch, UniPercepNet-S provides a practical solution for real-time perception. Its balance between simplicity, accuracy, and speed makes it especially valuable for autonomous driving and other embedded vision applications.
PurposeThe COVID-19 pandemic has imposed unprecedented strain on healthcare systems worldwide, largely due to significant and unpredictable fluctuations in patients' length of stay (LOS) during hospitalization. Interpretable deep learning techniques have emerged as promising tools for accurately predicting LOS and thereby optimizing service levels and resource allocation. Design/methodology/approachIn this study, we extracted electronic health record (EHR) data from a tertiary hospital in Wuhan for COVID-19 patients, who were stratified into three groups according to their LOS: short (<3 days), intermediate (3–5 days) and long (>5 days). We applied nine predictive algorithms, including three conventional machine learning methods – logistic regression (LR), support vector classification (SVC) and decision tree (DT) – and three deep learning models – AutoInt (AU), xDeepFM (XD) and FiBiNet (FB). To enhance model transparency and clinical applicability, we employed Shapley Additive Explanations (SHAP) to interpret model outputs and identify key variables influencing LOS. FindingsOur results revealed several clinical and demographic variables significantly associated with LOS. Among all tested models, the FiBiNet (FB) model achieved the best predictive performance, with an average improvement of approximately 3% over other models. Compared with traditional machine learning methods, deep learning models demonstrated an average performance gain of 1.7%. Interpretable analysis further revealed that medical insurance status was a crucial determinant of LOS, particularly in differentiating between patients with household and non-household registration in Wuhan. Originality/valueTo the best of the authors’ knowledge, this study is among the first to integrate interpretable deep learning methods with EHR data to predict LOS for COVID-19 patients in Wuhan. By identifying medical insurance status and self-financing as key predictors, our findings offer actionable insights for hospital administrators seeking to optimize medical resource allocation. The results contribute to strengthening China's healthcare service system and support ongoing efforts in medical and healthcare reform.
PurposeAccurate user identification serves as the cornerstone for demand recognition in precision marketing. Previous studies have overlooked the distorting influence of disinformation in online environments. This study develops an innovative user positioning framework that explicitly accounts for disinformation interference.Design/methodology/approachThe proposed methodology employs a three-stage analytical process: First, advanced semantic analysis techniques identify and filter disinformation. Second, high-dimensional cluster analysis categorizes users through K-means optimization. Finally, a multidimensional integration of demographic attributes and behavioral preferences enables precise repositioning of high-impact users.FindingsThe framework demonstrates (1) significant improvement in disinformation detection accuracy through semantic analysis, (2) enhanced clustering purity in user segmentation, and (3) more reliable identification of high-impact users through multidimensional feature integration.Originality/valueThis study proposes a new method of high-impact user positioning by identifying the convoluting effect of disinformation on internet consumer communities.
PurposeThe purpose of this paper is to propose a hierarchical deep reinforcement learning (H-DRL) framework for real-time tactical decision-making in team sports. The framework addresses challenges such as continuous action spaces, partial observability and adversarial environments by leveraging multi-agent collaboration, adaptive strategy optimization and explainable artificial intelligence (AI). It aims to enhance tactical accuracy, decision speed and resource efficiency while uncovering novel tactical patterns that human experts may overlook.Design/methodology/approachThe study combines graph neural networks (GNNs) for spatial-temporal player interactions and transformer-based attention for strategic pattern recognition. It integrates opponent modeling via inverse reinforcement learning (IRL) and self-play. The hierarchical architecture decomposes decisions into strategic, tactical and technical levels. Experiments were conducted on professional basketball, soccer and rugby datasets, with validation by expert coaches. The framework was tested in real-world deployments, including youth academies and professional teams, to evaluate performance and tactical innovations.FindingsThe H-DRL framework achieved a 34.7% higher tactical accuracy, 28.3% faster decision-making and 41.2% lower resource usage compared to state-of-the-art methods. It identified 17 new tactical patterns, such as dynamic role-switching, which improved scoring efficiency by 23.6%. Real-world deployments demonstrated significant performance gains, including a 42.3% improvement in tactical decision-making for youth teams. The system's explainable AI module bridged algorithmic insights with coach expertise, fostering trust and adoption.Research limitations/implicationsThe study is limited by its reliance on proprietary tracking data and the computational demands of real-time deployment. Future research could explore cross-sport generalization and integration of physiological/psychological factors. The framework's scalability to larger team sizes and more complex environments remains a challenge. These limitations highlight opportunities for advancements in model compression and hardware optimization.Practical implicationsThe framework provides actionable insights for coaches and players, enhancing in-game decision-making and training efficiency. It enables teams to adopt data-driven tactics, such as elastic pressing in soccer or optimized phase play in rugby. The system's real-time capabilities (30.9 ms latency) make it suitable for live match analysis. Professional teams reported improved tactical understanding (91.7% of coaches) and scoring efficiency (23.6% increase). The technology is applicable beyond sports, including autonomous systems and emergency response.Social implicationsThe study promotes the ethical use of AI in sports, emphasizing augmentation over replacement of human expertise. It fosters collaboration between coaches and AI, enhancing tactical literacy and innovation. The framework's transparency builds trust, addressing concerns about black-box AI. By uncovering counterintuitive strategies, it challenges traditional coaching paradigms and encourages continuous learning. The technology's broader applications (e.g. military, robotics) underscore its societal impact.Originality/valueThis paper presents a hierarchical DRL framework for real-time tactical decision-making in team sports, integrating GNNs, transformers and IRL. Its dual-stream architecture and adaptive computation are novel contributions. The system's ability to discover and explain tactical innovations (e.g. dynamic role-switching) sets it apart from prior work. The rigorous validation across multiple sports and real-world deployments demonstrates its practical value. The study advances multi-agent AI, offering scalable solutions for complex, dynamic environments.
PurposeForecasting reservoir water levels plays a critical role in effective water resource management, contributing to the safety of hydraulic infrastructure and mitigating the impacts of droughts and floods. Current forecasting models work solely on satellite imagery, which has limitations on handling noise, particularly in peak values within the time-series streamflow from reservoir operation. Indeed, a multi-modal approach that integrates both satellite imagery and reservoir operation data is necessary to enhance the performance of forecasting.Design/methodology/approachThis research presents a novel multi-modal forecasting model that integrates satellite imagery with historical water level data to improve prediction accuracy, particularly in forecasting abrupt changes in water levels. Image features are extracted using the histogram of oriented gradients (HOG) algorithm and normalized with the L2 norm to enhance training stability and reduce noise. A customized fusion function is developed to combine spatial features from satellite imagery with temporal features from water level time series, resulting in a unified composite feature vector. This vector, along with the historical water level sequence, is fed into a gated recurrent unit (GRU) model for forecasting. The fusion mechanism plays a crucial role in capturing sudden and abnormal variations in the data.FindingsThe model is assessed using satellite images and on-site water level measurements collected at the An Khe and Ka Nak Reservoir, Gia Lai, Vietnam, spanning January 2019 to December 2022. Experimental results demonstrate that the HOG-GRU variant significantly outperforms conventional deep learning models. The specific evaluation metrics are as follows: for the An Khe Reservoir, mean squared error (MSE) (0.08060), root mean squared error (RMSE) (0.28390), mean absolute error (MAE) (0.20446) and |Tracking Signal| (0.00032); whereas for the Ka Nak Reservoir, the corresponding values are MSE (0.20795), RMSE (0.45601), MAE (0.37937) and |Tracking Signal| (0.03985). These findings confirm the model's robustness and its practical applicability to real-world hydrological forecasting tasks.Originality/valueThis paper presents an original research contribution, offering novel insights to the academic domain of information technology, with all references comprehensively and accurately cited.
Purpose One widely adopted approach for effectively handling class-imbalanced datasets is data over-sampling, which involves generating synthetic samples for the minority class. Among these methods, the synthetic minority oversampling technique (SMOTE) is one of the most commonly used algorithms across various domain-specific imbalanced datasets. Numerous SMOTE variants have been proposed in the literature to enhance performance. However, no single algorithm consistently outperforms others across all types of domain-specific datasets. Design/methodology/approach This paper introduces several ensemble over-sampling methods, which combine the results of multiple over-sampling algorithms applied in both parallel and sequential manners. In the experiments, 58 binary-class datasets are used in Study One, while Study Two involves 10 datasets from various medical domains, encompassing both binary and multi-class classification tasks. Additionally, four baseline over-sampling algorithms, i.e. SMOTE, Poly-Fit-SMOTE, ProWSyn and SMOTE-IPF, are evaluated alongside three classifiers: C4.5, SVM and XGBoost. Findings The experimental results demonstrate that the parallel-based ensemble method, which combines ProWSyn and SMOTE-IPF and selects the k nearest synthetic samples around the corresponding minority class centers, yields the best performance. Using this ensemble approach, the XGBoost classifier achieves superior AUC and G-mean results compared to those obtained with the four baseline over-sampling algorithms and most other ensemble methods. Originality/value This paper presents parallel and serial ensemble methods that integrate multiple over-sampling results to enhance the performance of individual over-sampling algorithms. The most effective ensemble method and its associated combined algorithms can serve as a representative baseline for future research in class imbalance learning.
Purpose Chinese government documents require sensitivity reviews to detect and protect sensitive personal information before public release. However, this process is often time-consuming and labor-intensive due to the lack of robust automated detection and classification systems, especially considering the complexities of Chinese morphology, grammar and semantics, which challenge existing solutions. Design/methodology/approach This study proposes a parallel dual-channel hybrid neural network model for the detection and classification of sensitive information. The model leverages Bidirectional Encoder Representations from Transformers (BERT) for text vector representation, combined with Text Convolutional Neural Networks (TextCNN) and Bi-directional Long Short-Term Memory (BiLSTM) networks to extract both local and global semantic features. The integration of an attention mechanism further enhances the model’s capacity by emphasizing critical terms, thereby improving the precision of sensitive information classification. Findings The effectiveness of the proposed model, CGDSClass, was validated through comparative analysis with nine existing models across two datasets. The results demonstrate that CGDSClass significantly outperforms these models, indicating a superior capability in handling the specific challenges of sensitive information detection in Chinese government documents. Originality/value As the first study to address sensitive information detection and classification in the context of Chinese government open data, this research offers a novel theoretical and technical foundation. CGDSClass not only improves the accuracy of classification but also enhances early warning and monitoring capabilities, thereby reducing the risk of data breaches and advancing the openness of government data practices.
Purpose The purpose of the proposed work is to enhance Arabic-English translation and other Arabic NLP tasks by addressing the unique linguistic challenges of the Arabic language. The Arabic Language Attention Transformer (ALAT) integrates morphological awareness, diacritic sensitivity, and right-to-left positional encoding to improve semantic understanding, syntactic accuracy, and contextual representation. By overcoming limitations in existing models like AraBERT and mBERT, ALAT aims to deliver superior performance in sentiment analysis, machine translation, and named entity recognition, setting a new benchmark for Arabic NLP applications. Design/methodology/approach The proposed ALAT employs an advanced attention-based architecture tailored for Arabic-English translation. It integrates a Morphological Awareness Module (MAM) to capture root-based patterns, a Diacritic-Aware Embedding Layer for semantic disambiguation, and a custom Right-to-Left Positional Encoding aligned with Arabic script. The model enhances the standard Transformer by incorporating morphological and diacritic features into its self-attention mechanism, improving contextual representation. ALAT is trained on diverse Arabic-English parallel corpora using cross-entropy loss with optimization techniques like the Adam optimizer, achieving superior performance in translation accuracy, as demonstrated by improved BLEU scores. Findings The proposed ALAT model significantly outperforms existing Arabic NLP models in key tasks. It achieved an F1-score of 92.4% in sentiment analysis, 86.1% in named entity recognition, and a BLEU score of 47.8 in machine translation, surpassing models like AraBERT and mBERT. ALAT's Morphological Awareness Module and Diacritic-Aware Embedding Layer improved semantic accuracy, while right-to-left positional encoding enhanced syntactic coherence. The model also demonstrated superior performance in morphological disambiguation (94.5% accuracy) and diacritic restoration (97.2% accuracy), establishing its effectiveness in addressing Arabic's linguistic complexities. Originality/value The originality of the proposed work lies in the development of the ALAT, a model specifically designed to address Arabic's unique linguistic challenges. Unlike existing models, ALAT integrates a Morphological Awareness Module, Diacritic-Aware Embedding Layer, and Right-to-Left Positional Encoding, enhancing semantic understanding and syntactic accuracy. Its innovative attention mechanism captures both morphological and diacritic features, significantly improving performance across NLP tasks. This work adds substantial value by setting a new standard for Arabic NLP, offering a robust, language-specific solution that outperforms general multilingual models in translation, sentiment analysis, and named entity recognition.
Purpose This study explores the pervasive gender bias in the editorial decision-making processes of several Wikipedia language editions (Catalan, Italian and French). It examines how these processes – focused on deliberations and voting – affect the representation of gender diversity. Specifically, it evaluates the implications of banning gender-based categories for accessing information, rejecting inclusive language and the controversial practice of including trans individuals' dead names in biographies. Design/methodology/approach The research employs a detailed content analysis of arguments presented during nine deliberative and voting processes conducted between 2014 and 2024. A redesigned codebook was utilised to classify and analyse arguments based on their type and stance toward gender inclusion. The study also coded editors' gender identities, where identifiable, to assess patterns of participation and perspectives within these processes. Findings The findings reveal a systemic resistance to gender diversity in decision-making processes, predominantly justified through appeals to values, neutrality policies and community norms. While arguments supporting inclusivity often emphasised ethical considerations, practical reasoning and precedence, they were insufficient to sway outcomes in favour of gender inclusion. The lack of consensus and resistance perpetuates the invisibility of gender-diverse identities, reinforcing existing inequalities. Originality/value This research uniquely examines Wikipedia's decision-making infrastructure and its broader implications for knowledge representation. By highlighting the systemic barriers to inclusivity in collaborative knowledge platforms, it contributes to critical discussions on diversity and equity in digital information systems.
PurposeThe global rise of alternative medicine, particularly the use of medicinal plants, offers promising health benefits but also poses significant risks due to limited knowledge, species misidentification and improper preparation or administration.Design/methodology/approachThis paper introduces HerbOnto, a Herbal Ontology designed for the dynamic development of medicinal plant knowledge base - meaning it continuously collects, updates and categorizes data through advanced data collection techniques and classification algorithms.FindingsHerbOnto aims to reduce risks and enhance safety by providing an evolving, structured and accessible knowledge base that supports the effective and responsible use of medicinal plants worldwide.Originality/valueThe dynamic and predictive approach to ontology development for medicinal plants, which goes beyond existing static herbal databases or taxonomies.
