Forthcoming Articles
International Journal of Information and Communication Technology
Forthcoming articles have been peer-reviewed and accepted for publication but are pending final changes, are not yet published and may not appear here in their final order of publication until they are assigned to issues. Therefore, the content conforms to our standards but the presentation (e.g. typesetting and proof-reading) is not necessarily up to the Inderscience standard. Additionally, titles, authors, abstracts and keywords may change before publication. Articles will not be published until the final proofs are validated by their authors.
Forthcoming articles must be purchased for the purposes of research, teaching and private study only. These articles can be cited using the expression "in press". For example: Smith, J. (in press). Article Title. Journal Title.
Articles marked with this shopping trolley icon are available for purchase - click on the icon to send an email request to purchase.
Online First articles are also listed here. Online First articles are fully citeable, complete with a DOI. They can be cited, read, and downloaded. Online First articles are published as Open Access (OA) articles to make the latest research available as early as possible.
Articles marked with this Open Access icon are Online First articles. They are freely available and openly accessible to all without any restriction except the ones stated in their respective CC licenses.Register for our alerting service, which notifies you by email when new issues are published online.
International Journal of Information and Communication Technology (32 papers in press) Regular Issues
Abstract: In order to meet the demand for the digital modelling and intelligent analysis of students learning behaviour, a predictive evaluation framework is proposed, which combines feature engineering with LightGBM. Through the construction of tags and the extraction of training samples from college English course behaviour log, the construction strategy and interpretable output mechanism of the model are discussed. The results show that the proposed model achieves an F1-score of 0.889 on the test set, with an average response latency of less than 230 ms, and a performance fluctuation of less than 1.6% in 5 times cross validation, demonstrating robust deployment stability and generalisation. Keywords: learning behaviour prediction; LightGBM algorithm; educational data mining; Intelligent teaching analytics; learning behaviour modelling. DOI: 10.1504/IJICT.2027.10080149
Abstract: To improve the accuracy of key indicator evaluation in power engineering, this study proposes a comprehensive evaluation method based on K-nearest neighbour (KNN) algorithm. Aiming at the limitations of traditional methods in processing nonlinear and high-dimensional data, an evaluation system including ten key indicators such as engineering cost, construction period, and equipment performance was constructed to conduct experiments on 50 power engineering project data in a certain region. The results show that compared with traditional methods, the KNN model has a 15% improvement in evaluation accuracy, a more reasonable allocation of indicator weights, and can effectively identify key factors affecting engineering quality and efficiency. It also exhibits strong robustness in dealing with nonlinear problems and is suitable for power engineering of different scales and types. This method provides a scientific basis for decision-making in power engineering and has promotional value. Keywords: KNN algorithm; power engineering; index analysis; comprehensive evaluation. DOI: 10.1504/IJICT.2026.10080314 An international communication influence model for intangible cultural heritage using graph transformers and MoE architecture by Jing Zeng Abstract: International dissemination of intangible cultural heritage (ICH) is difficult to assess because relevant data are heterogeneous, relational, and dynamically evolving. We propose GT-MoE-NIA, an influence assessment model that combines a graph transformer with a mixture-of-experts (MoE) architecture. A heterogeneous graph represents ICH projects, countries or regions, media, audiences, and their interactions. The graph transformer captures global structural and semantic dependencies, while a gating network dynamically routes representations to specialised experts for influence prediction. We evaluate the model on the Global Non-Material Cultural Heritage Propagation Dataset (GNCPD), integrating UNESCO inventories, GDELT news, Twitter data, and Wikipedia page views. Compared with GCN, GAT, and GraphSAGE baselines, GT-MoE-NIA reduces root mean square error by 18.7% and improves NDCG@10 by 22.4%. Expert-activation patterns also provide interpretable evidence of how dissemination modes affect influence, supporting data-driven resource allocation for international cultural promotion. Keywords: intangible cultural heritage; ICH; international dissemination; influence model; graph transformer; mixture of experts; heterogeneous graph neural network.  This article is being considered for Open Access
Abstract: This study proposes a multi-agent cross-language cognitive computing framework grounded in cognitive translation studies to enhance humanmachine collaborative translation. The framework integrates specialised agents for drafting, contextual reasoning, knowledge retrieval, revision, and evaluation, aiming to address hallucinations and logical inconsistencies in large language model-based translation. A computational simulation involving 50 virtual professional translators was conducted to evaluate translation quality and simulated cognitive workload. The framework achieved COMET-22 scores of 0.871 in general translation and 0.846 in biomedical translation, with approximately 85% fewer factual hallucinations in the simulated environment. The simulated key stroke correction rate decreased to 6.2%, alongside reductions in simulated gaze time and peak pupil dilation. These findings suggest potential benefits for translation quality and cognitive offloading. All translator behaviours, eye-tracking measurements, keystroke dynamics, and cognitive-load data were synthetically generated; no human participants were involved. Keywords: cognitive translation studies; large language models; LLMs; human-in-the-loop translation; multi-agent cognitive computing; cognitive load; factuality hallucination. DOI: 10.1504/IJICT.2026.10082146
Abstract: Coaches divide aerobics stages visually, which is slow and inconsistent. Deep-learning models segment actions well but are black boxes needing large labeled sets. We propose an interpretable machine-learning framework, CGTTS, that extracts center-of-gravity kinematic featuresvelocity, acceleration, speed magnitudeand learns a velocity-acceleration joint criterion (VAJC) to segment continuous aerobics sequences into biomechanically meaningful units. Boundaries are set at velocity extrema and acceleration zero-crossings, giving physically interpretable decisions. Cast as frame-level boundary classification and validated by five-fold cross-validation on 30 subjects, VAJC reaches 0.882 accuracy, beating force-time peak detection by 14.8 points and the deep-learning MS-TCRNet baseline on trajectory accuracy (0.842) and phase consistency (0.803), gains of 8.1% and 10.5%. The framework offers a data-efficient, transparent tool that cuts subjective teaching variability. Keywords: temporal motion segmentation; velocity-acceleration criterion; interpretable machine learning; biomechanical analysis; aerobics instruction. DOI: 10.1504/IJICT.2026.10082172
Abstract: In standardised tests and online education, math question supply still relies on manual creation, which is inefficient and fails to meet the differentiated needs of learners. Existing automated methods treat difficulty as a post-generation attribute rather than a pre-specified condition, making on-demand generation difficult. Difficulty ranking also lacks objective quantification, and manual annotation introduces subjectivity that hinders intelligent question bank management. To address these issues, this paper proposes diffusion-based math item generation, which uses difficulty coefficients and knowledge point labels as explicit conditions, synthesises question text via gradual denoising in continuous space, and incorporates item response theory for quantitative difficulty estimation and ranking. Experiments show that diffusion-based math item generation achieves a difficulty mean absolute error of 0.087, 59.3% lower than diffusion models without explicit difficulty modelling, and a ranking consistency of 0.912, 24.1% higher than the best baseline, confirming the effectiveness of embedding difficulty priors into diffusion generation. Keywords: diffusion model; controllable text generation; test difficulty; item response theory; IRT. DOI: 10.1504/IJICT.2026.10082173
Abstract: When AI aesthetic feedback is introduced into the hand-eye interaction loop of mobile painting, it can perturb the users coordination patterns and drive their reshaping, but lacks quantitative characterisation and optimisation methods. To address this, this paper first incorporates hand-eye movements into an eight-dimensional state space, quantifies baseline coupling with mutual information, and defines instantaneous perturbation with normalised state distance. Then, a Bayesian dynamics model is constructed to recursively estimate the internal aesthetic learning rate using Kalman filtering, and a reshaping index is defined to discriminate coordination structural changes. Furthermore, priority replay and adaptive exploration noise are used to improve the deep deterministic policy gradient algorithm, jointly optimising feedback modality and update timing. Experiments imply that the suggested approach reduces instantaneous perturbation intensity by 38.2% compared to the best baseline and improves reshaping level by 13.2%, verifying the effectiveness of the suggested method. Keywords: mobile painting; AI aesthetic feedback; hand-eye interaction; perturbation and reshaping; reinforcement learning. DOI: 10.1504/IJICT.2026.10082174 Cross-domain secure sharing mechanism for government databased on federated learning and blockchain by Xianwei Chen, Lixiang Zhou, Min Zhang Abstract: Cross-domain government data sharing faces an inherent tension between the requirement that sensitive data must not leave its domain and the strong demand for cross-departmental collaboration driven by heterogeneity. Traditional centralised schemes struggle to simultaneously satisfy privacy protection, feature alignment, and process auditability in untrusted environments. To address this, we propose a secure sharing model integrating federated learning, differential privacy, blockchain, and graph convolutional networks. Local federated training blocks direct raw data aggregation, while differential privacy noise quantitatively resists parameter reconstruction attacks. Government departments are mapped to graph nodes with business interaction strength as edge weights, and graph convolutional networks extract implicit cross-departmental dependencies to alleviate heterogeneous feature alignment difficulties. Blockchain-calculated client confidence serves as the trusted weight for federated aggregation, replacing sample-size-based weighting. Experiments show 97.53% average accuracy, 0.198 ms average transmission time, and 99.50% decryption success, significantly outperforming comparison models. Keywords: federated learning; FL; blockchain technology; BT; data sharing; cross-domain security; differential privacy; DP; graph convolutional network; GCN. This article is being considered for Open Access
Abstract: Traditional dance preservation in Shanxi faces generational gaps, cultural deterioration, incomplete documentation of intangible cultural assets, and limited digital reconstruction tools. To tackle these challenges, this study proposes an artificial intelligence-based deep heritage reconstruction model (AI-DHRM) for reconstructing the deep structure of Shanxi traditional dances. The system extracts the stylistic features of the historical dances, finds the hidden movement features, and maps the cultural meanings to the semantic dance features. It combines motion pattern recognition, deep neural feature extraction, cultural semantic mapping, and adaptive sequence learning to uncover hidden stylistic elements. AI-DHRM enhances dance interpretability and reconstructs original structures using multimodal inputs, such as motion-capture data, archived films, and ethnographic annotations. The model achieved a 95.85% recall, 97.08% reconstruction accuracy (RA), and a 95.23% semantic preservation score (SPS). AI-DHRM provides a solution to the preservation, restoration, and transmission of the intangible dance heritage of Shanxi, while retaining symbolism, rhythm, cultural semantics, and continuity. Keywords: artificial intelligence; AI; intangible cultural heritage; ICH; Shanxi dance traditions; deep learning; cultural reconstruction; motion pattern recognition; digital heritage preservation; AI-DHRM. DOI: 10.1504/IJICT.2026.10082242
Abstract: Urban rail transit (URT) systems face critical evacuation challenges during emergencies due to high density, complex layouts, and dynamic crowd behaviour. This paper presents an IoT and edge computing framework for real-time crowd evacuation optimisation, integrating intelligent sensing, adaptive decision making, and low-latency responses. Using the URT Evacuation Behaviour Dataset comprising CCTV, heat/smoke sensors, Wi-Fi/RFID logs, and simulated paths data are pre-processed via resizing, Butterworth filtering, cleaning, normalisation, and synchronisation. A MobileNetV2-based CNN estimates crowd density, while the APPFN model (autoencoder for anomaly detection, graph neural network for flow prediction, and proximal policy optimisation for adaptive routing) generates optimal evacuation strategies. MATLAB simulations based on the URT Evacuation Behaviour Dataset and station-level scenario settings achieve a benchmark evacuation time of 260 s in the comparative evaluation and scenario-specific evacuation times of 340388 s across the tested emergency scenarios, with a mean absolute error of 22.98, demonstrating significant improvements over conventional methods. Keywords: crowd evacuation; decision making; smart transit infrastructures; routes; detection; graph neural network; GNN. DOI: 10.1504/IJICT.2026.10082251
Abstract: Traditional intrusion detection systems and black-box deep learning (DL) models often fail to provide transparent reasoning behind threat classification, thereby limiting trust, interpretability, and incident response efficiency in cybersecurity operations. To address these challenges, this research proposes an explainable artificial intelligence (XAI)-driven DL framework to interpret decision pathways in APT detection workflows. The proposed approach integrates a white shark optimiser-driven deep neural network with an attention-mechanism (WSO-DNN-AM) for behavioural analysis to identify and interpret malicious network activities from high-dimensional traffic data. The proposed approach performs multilayer analysis of abnormal connections, traffic behaviour profiles, and intrusion patterns to detect stealthy APT activities. On the APT detection traffic analytics dataset, the model achieved 98.65% detection accuracy and outperformed the compared conventional methods under the same experimental protocol. Furthermore, the interpretability modules provide transparent reasoning for threat predictions, enabling cybersecurity analysts to understand critical attack indicators, improve decision-making, and strengthen proactive cyber defense mechanisms in advanced enterprise security environments. Keywords: artificial intelligence; advanced persistent threat detection; deep neural network; cyber threat intelligence; network traffic analysis. DOI: 10.1504/IJICT.2026.10082252
Abstract: Employee turnover is shaped by individual behavioural trajectories and organisational relationship structures. Existing models treat employees independently, neglecting peer influence through collaboration networks, while static approaches fail to capture the gradual evolution of turnover intention. This paper proposes the graph attention-bidirectional long short-term memory for attrition prediction model, which encodes management reporting and project collaboration ties as a graph structure. The graph attention network captures structural positions, and the bidirectional long short-term memory network models behavioural dynamics including satisfaction and performance over time, enabling joint feature interaction. Experiments on four-year anonymised data from 8,326 employees show that the model achieves an area under the curve of 0.957, outperforming the transformer baseline by 2.0%, and attains a dynamic warning timeliness of 3.4 quarters, surpassing the baseline by 0.6 quarters, confirming its ability to issue early warnings nearly one year before departure. Keywords: employee attrition prediction; graph attention network; GAT; bidirectional long short-term memory; dynamic risk modelling. DOI: 10.1504/IJICT.2026.10082253
Abstract: Can policies promoting industry-education integration truly improve corporate productivity? Non-random selection in pilot programs often leads to evaluation bias. This paper introduces a latent outcome framework and balanced representation learning, using neural networks to map companies multidimensional characteristics into comparable abstract representations, thereby approximating randomised controlled conditions in the observed data and quantifying the causal effects of the policies. Using the 2021 national pilot program as a quasi-experiment and drawing on data from A-share companies from 2016 to 2024, the model estimates that the policy increased total factor productivity by an average of 11.3%, nearly 30% higher than the 8.7% estimated by traditional difference-in-differences methods; when identifying firms positively affected by the policy, the area under the curve reached 0.874, outperforming the 0.791 achieved by causal forests. The results indicate that, after correcting for selection bias, the policys actual effectiveness is more pronounced than that shown by traditional methods. Keywords: industry-education integration; causal effects; potential outcomes; representation learning; policy evaluation. DOI: 10.1504/IJICT.2026.10082254
Abstract: League of Nations technical cooperation with China produced reports, maps, inventories, and correspondence that form essential evidence for studying modern international aid. Yet scattered storage and semantic gaps across modalities have long constrained systematic analysis. This paper proposes dual-modal semantic fusion, which introduces adaptive fusion weights to balance text and visual features in a shared space and leverages graph structures to discover cross-modal associations. On the CuLi test split containing 5,726 image-text pairs (7:1:2 split), the framework achieves an area under the curve of 0.876, a 6.3-percentage-point improvement over the strongest baseline language-aligned Chinese contrastive language-image pre-training, alongside a semantic consistency score of 0.743 and a cross-modal link density of 0.612, indicating over sixty percent of graph connections bridge text and images. Adaptive weighting improves the area under the curve from 0.835 to 0.876 over equal-weight fusion, filling the gap of dedicated cross-modal fusion for historical archives. Keywords: cross-modal fusion; historical archives; semantic consistency; linkage discovery. DOI: 10.1504/IJICT.2026.10082255
Abstract: In supply chain operations, dynamic disturbances such as demand fluctuations, transportation disruptions, and carbon price changes persist. Traditional scheduling models treat emissions as a rigid constraint rather than an optimisation objective, failing to balance cost and emissions dynamically. Existing methods lack real-time adaptability under frequent disturbances. This paper proposes the carbon-aware hierarchical reinforcement learning framework, which decomposes decisions into strategic and tactical layers, and introduces a Pareto-optimal reward mechanism with adaptive weights to resolve multi-objective conflicts. Experiments on public logistics datasets show that Compared to the optimal hierarchical reinforcement learning baseline, the framework reduces total cost by 6.0% and carbon emission intensity by 6.5%. Against the single-layer deep reinforcement learning baseline, the on-time delivery rate rises by 1.9 percentage points and the scheduling stability index falls by 20.5%. This framework bridges carbon-cost optimisation and real-time response in dynamic scheduling, offering a feasible path for intelligent scheduling under carbon constraints. Keywords: supply chain scheduling; deep reinforcement learning; DRL; carbon emission reduction; multi-objective optimisation. DOI: 10.1504/IJICT.2026.10082256
Abstract: The degree to which course objectives are met serves as a key basis for engineering education accreditation and program evaluation; however, existing prediction methods overlook the topological dependencies among knowledge points and rely on final exams, resulting in delays. To address this, we introduce knowledge graphs and random walks to propose a dual-mode walking prediction framework: we construct a course graph based on the syllabus and assessments; capture hierarchical relationships and functional similarities through structural and semantic walks; and map sparse behavioural data to a variable space. Validated using data from three courses involving over 4,286 students, this method achieved an accuracy of 88.7% and an area under the curve of 0.956, representing improvements of 2.5 and 2.2 percentage points, respectively, over the graph attention network baseline. The results confirm that knowledge graph-based random walks can effectively mitigate data sparsity and enhance interpretability. Keywords: knowledge graph; random walk; course objective achievement; learning prediction. DOI: 10.1504/IJICT.2026.10082257
Abstract: Emotional resonance in values education lies in speech, text, and facial expressions. Due to the multimodal semantic gap, existing models fail to capture the awakening and migration of value identification from cross-modal signals, and lack a computable evolutionary framework. This paper proposes the emotional resonance computation evolutionary graph, which maps teaching events to graph nodes and encodes resonance states as edge trajectories via cross-modal alignment and state transition regularisation, enabling structured tracking of the awakening-reinforcement-internalisation process. Experiments show accuracy improves from 75.2% to 78.6% and area under the curve from 0.817 to 0.841 over the strongest baseline self-supervised multi-task multimodal. On CMU-MOSEI, the emotional resonance coefficient rises from 0.62 to 0.74, and value identification consistency increases by 33.3% over the best baseline. These results demonstrate the framework fills the gap in computable evolutionary modelling of emotional resonance for values education, offering a quantitative tool for classroom assessment. Keywords: emotional resonance; semantic alignment; evolutionary graph; multimodal learning. DOI: 10.1504/IJICT.2026.10082244
Abstract: Logical coherence is crucial for the readability and quality in academic writing, but the traditional approaches to coherence analysis are not well suited to the analysis of complex semantic and discourse-level relationships. In this work, an LSTM-Attention model is introduced to address the efficient capturing of discourse transitions which combines the sequential memory network with adaptive sentence-importance weighting. The proposed method makes emphasis on progressive discourse relationship, while BiLSTM-Attention models attention following the bi-directional contextual encoding. It is more computational efficient than Transformer-based models, and offers greater interpretability. Pre-processing of curated essays and benchmark corpus is performed using standard NLP techniques: tokenisation, sentence segmentation, normalisation and GloVe embeddings with human coherence scores. The MAE, MSE, Pearson correlation, Spearman correlation and R2 of the proposed model are 0.95, 0.038, 0.91, 0.90 and 0.90 respectively. The results are very consistent with human judgments and significantly better than benchmark baseline methods. Keywords: logical coherence; LSTM; attention mechanism; English writing analysis; DL-based text evaluation. DOI: 10.1504/IJICT.2026.10081884 Abstract: Generative AI systems in high-stakes risk management frequently suffer from reward overoptimisation, hallucinations, and adversarial prompt vulnerabilities, undermining decision reliability. To mitigate these risks, an adversarially robust prompt tuning framework with reinforcement learning from human feedback (ARP-RLHF) is introduced for secure and human-aligned decision support. A prompt risk dataset of 10,000 records covering healthcare, finance, and cybersecurity is pre-processed with text normalisation, tokenisation, and TF-IDF feature extraction. Unlike existing RLHF and prompt tuning methods, ARP-RLHF jointly optimises adversarial prompt tuning, enriched proximal policy optimisation, and the reptile search algorithm to strengthen adversarial resilience, policy learning, and parameter optimisation. Experimental results demonstrate that ARP-RLHF achieves 95%-97% accuracy and outperforms existing methods, delivering enhanced safety compliance, robustness, and hallucination reduction for trustworthy risk management decisions. Keywords: reinforcement learning from human feedback; RLHF; prompt tuning; AI risk mitigation; high-stakes decision systems; trustworthy. DOI: 10.1504/IJICT.2026.10082043
Abstract: The rapid expansion of algorithmic trading (AT) has transformed global financial markets; however, its downstream effects on household consumption remain underexplored. Traditional econometric models often fail to capture AT's high-frequency, nonlinear impacts and are unable to integrate micro-level consumption data with real-time trading metrics. This research develops a machine learning augmented econometric framework to quantify the effects of AT on household spending. Data were collected from household expenditure surveys, high-frequency AT market microstructure datasets, and macro-financial indicators. Comparative evaluation demonstrates superior predictive performance, with an accuracy of 99.42%, a precision of 98.73%, a recall of 97.59%, an F1-score of 95.41%, a G-Mean is 98.03% and an AUC-ROC is 97.09%, outperforming conventional ML and deep learning benchmarks. Results reveal that higher AT activity significantly influences consumption via wealth and credit channels, particularly among higher-income and older households, with effects amplified during monetary tightening and volatile market regimes. Keywords: algorithmic trading; AT; enhanced golden eagle-based dynamic categorical boosting; EGE-Dyn-CatBoost; machine learning; ML; financial econometrics. DOI: 10.1504/IJICT.2026.10082129
Abstract: Centralised analytics and post-processing restrict real-time decision making, while wearable sensor networks provide continuous monitoring of athlete performance, physiology and biomechanics. We propose an agentic AI system that combines multimodal wearable IMU and HR and GPS sensors with edge computing. Streaming data is locally analysed with light deep learning models and reinforcement learning agents to evaluate fatigue, efficiency of movement, workload, and risk of injury, and to give adaptive prompts during training. The framework has been validated from 120,000 synchronised sensor samples from 50 athletes running endurance, football drills and interval training in four states (normal, fatigue, optimal performance and injury risk). Edge inference, implemented in Python, TensorFlow Lite and PyTorch, is tested on the Raspberry Pi 4. Experimental results show that the accuracy of the system is 96.4%, the precision is 96.0%, the recall is 95.8%, the decision latency is 38 ms, and the energy consumption is 10.0 J. Keywords: agentic artificial intelligence; wearable sensor networks; sports analytics; edge computing; autonomous decision system. DOI: 10.1504/IJICT.2026.10081885
Abstract: The driving effects of the digital economy on the manufacturing sector vary significantly across different dimensions, yet existing research lacks a systematic quantification and ranking of the contributions of each dimension. This paper constructs a knowledge graph of the integration between the digital economy and manufacturing and proposes an association ranking framework based on a graph attention network to address the challenge of comparing correlation strengths across multi-source, heterogeneous data. Experiments were conducted using provincial-level panel data from China for the period 2010-2024. The results indicate that the marketisation of data factors and industrial digitisation are the core dimensions driving manufacturing upgrading, with an area under the curve of 0.862 - a 5.9 percentage point improvement over the baseline graph convolutional network - and a normalised discounted cumulative gain @10 of 0.758, representing a 4.6 percentage point improvement over the best baseline model. Keywords: knowledge graph; digital economy; manufacturing upgrading; graph attention network; association ranking. DOI: 10.1504/IJICT.2026.10082044
Abstract: Conventional phantom-based ultrasound transducer quality control relies on subjective image interpretation and may miss early element-level degradation. This study developed an intelligent framework integrating channel-level acoustic measurements, statistical feature fusion, and machine learning for automated transducer assessment. A total of 113 clinically used transducers underwent phantom testing and acoustic measurement. Six acoustic parameters were converted into 30 probe-level statistical features, and five classifiers were evaluated using three-fold cross-validation. Phantom indicators showed no significant group differences, whereas 15 acoustic features differed significantly, with four remaining significant after multiple-comparison correction. Random forest achieved balanced performance (accuracy 96.26%, F1-score 90.59%, AUC 0.9878) with interpretable feature importance. Sensitivity variation was the most informative feature, suggesting that channel-to-channel dispersion may indicate degradation earlier than changes in mean values. These findings support acoustic sensing as a quantitative complement to phantom testing, although external multi-centre and longitudinal validation remains necessary. Keywords: intelligent acoustic sensing; ultrasound transducer; interpretable machine learning; equipment health assessment; channel-level signal processing; predictive maintenance; automated quality control. DOI: 10.1504/IJICT.2026.10082109
Abstract: Neural machine translation often exhibits inconsistencies or errors in the handling of specialised terminology. While large models offer superior performance, their high computational cost makes them difficult to deploy in practical settings. To address the issue of knowledge loss caused by model compression, this paper proposes a method that combines progressive distillation with terminology-aware attention, enabling small student models to learn sequentially from teachers of varying scales and to utilise bilingual terminology alignment information during training. Experiments in English-to-Chinese translation across three domains - medicine, law, and technology - show that, compared to the baseline, improved by 4.4-4.6 bilingual evaluation understudy points, the model size is only 5.6% of the largest teacher model (a 94.4% reduction), and the gap with the best teacher model was only 0.6 bilingual evaluation understudy points. The results indicate that this distillation strategy can effectively transfer terminology knowledge to lightweight models, facilitating deployment in resource-constrained environments. Keywords: neural machine translation; NMT; knowledge distillation; terminology translation; English-to-Chinese translation; model compression. DOI: 10.1504/IJICT.2026.10082045
Abstract: Current multimodal football action recognition faces challenges in fusing inertial and visual signals, adaptively weighting modalities, and recognising complex movements under visual occlusion. This study proposes a multimodal system combining inertial sensors and computer vision to detect football actions and evaluate long-term motion patterns. The framework combines dynamic time warping, spherical coordinate mapping, quaternion-based posture representation and temporal sequence classification. Angular velocity features are extracted with a two-layer recurrent neural network, and visual attention weights are generated based on the keypoint confidence and optical flow difference. A graph neural network models spatial-temporal dependencies for multimodal classification and alignment. Trajectory updating and smoothing using biomechanical joint constraints and a probabilistic field model. The framework combines adaptive weighting, spatiotemporal alignment, biomechanical constraints and long-term pattern analysis. In 20 experiments with 30 professional male football players, we achieved 88.8%, 0.89, and 81.4% for classification, clustering, and recognition under 50% camera occlusion. Keywords: football action recognition; inertial motion sensors; spatiotemporal alignment; biomechanical modelling; pattern clustering; visual-inertial integration; intelligent sports training. DOI: 10.1504/IJICT.2026.10082130
Abstract: Environmental, social and governance (ESG) reporting is becoming a key element in evaluating and informing sustainable investment choices. However, it is difficult to obtain structured ESG metrics from corporate reports because of the lack of consistency in format, unstructured narratives, and reporting practices. This paper outlines a novel multimodal ESG extraction and synthesis framework that leverages a red panda optimised attention deep graph convolutional neural network (RPO-ADGCNN), along with multimodal large language models (LLMs) and validation procedures. Experimental results on Python 3.11 show that the proposed approach achieves 92% accuracy, 0.18 MAE, and 0.22 RMSE on the corporate ESG metrics dataset, outperforming conventional models. The proposed model enables effective, scalable, and robust automated ESG analytics, facilitating transparent reporting by companies, regulatory adherence, and sustainable investment decisions. Keywords: environmental, social and governance; ESG; multimodal large language models; investment decision-making; sustainability targets. DOI: 10.1504/IJICT.2026.10082041
Abstract: In cross-cultural content distribution, narrative sequence significantly influences audience reception, yet existing systems overlook the interaction between sequence and cultural schemas. This paper proposes the culture-aware narrative sequence preference model, which uses cultural features as conditions, employs attention-based decoding to predict sequence scores, and searches for the optimal arrangement. Based on data from East Asia and Europe and the USA, the model was evaluated using area under the curve, accuracy, recall, and normalised discounted cumulative gain @K. The model achieved an area under the curve of 0.873, and an 8.1 percentage point improvement over the 0.792 achieved by the culture-agnostic model. Its normalised discounted cumulative gain @K (5th position) was 0.641, outperforming neural collaborative filtering (0.456) and the random model (0.251). The results indicate that the interaction between sequence and cultural schemas cannot be ignored, and this framework provides a computable basis for ranking. Keywords: narrative order; cross-cultural reception; perception of preferences; sequence modelling; cultural dimensions. DOI: 10.1504/IJICT.2026.10082046
Abstract: The risks in engineering projects are woven into a complex and dynamic network through process logic and causal relationships. Existing methods ignore the structured propagation and cascading effects of risks. Hence, this article suggests a temporal knowledge graph risk-reasoning model for project management. By introducing fuzzy time annotations and Bayesian completion mechanisms, the fuzzy expressions are modelled as probability distributions to restore the objective timeline. A time-aware hierarchical graph attention encoding for causal paths is proposed. Then, a risk cascading propagation equation with structural entropy regularisation and adaptive gating is constructed, allowing the propagation intensity to be dynamically adjusted according to the project stage. In addition, causal reasoning is used to convert predictions into decision support. Experimental outcome implies that the F1-score of the proposed model reaches 0.918, and the reasoning time is 4.7 seconds, outperforming benchmark methods, opening up new ideas for the intelligent transformation of project risk management. Keywords: project management; temporal knowledge graph; risk prediction; causal reasoning; graph attention. DOI: 10.1504/IJICT.2026.10082048
Abstract: Short-term voltage stability (STVS) in modern distribution systems is challenged by high penetration of distributed energy resources (DERs) such as PV, ESS, EVs, and demand response. This paper proposes an evolutionary-based game theoretic strategy that integrates the Ebola optimisation search algorithm (EOSA-EGT). DERs are modelled as autonomous agents in a multi-agent system; each agent selects between support and non-support strategies, with payoffs evolving via replicator dynamics to achieve equilibrium between economic efficiency and voltage stability. EOSA optimises these payoff functions, enhances convergence, and avoids local optima, while the objective minimises the inverse of the total voltage stability factor under operational constraints. Validated on the IEEE 36-bus system under dynamic operating conditions, the approach achieves 8.95 kW minimum power loss, 95.58% power loss reduction, and 44.32% voltage deviation reduction, providing an effective framework for intelligent energy management in smart distribution networks. Keywords: short-term voltage stability; evolutionary game theory; EGT; multi-agent systems; MAS; distributed energy resources; DERs. DOI: 10.1504/IJICT.2026.10082042
Abstract: In blended learning, the disconnection between online and offline behavioural data poses dual challenges of data sparsity and insufficient semantic utilisation for cognitive diagnosis. Existing neural cognitive diagnosis methods rely on expert-annotated Q matrices for concept modelling, failing to uncover implicit semantic associations in test question texts. While graph structure modelling captures interaction relationships, it lacks content understanding. This paper proposes the semantic-structure collaborative cognitive diagnosis model, which leverages semantic descriptions of test questions and student cognitive portraits generated by large language models as enhancement signals, injecting them into a heterogeneous graph constructed by graph neural networks to jointly model knowledge concept associations and student-test interaction patterns. Experiments on a public educational dataset demonstrate that the model achieves an area under the curve of 0.796, outperforming the optimal baseline by 1.7 percentage points, and attains a cognitive consistency index of 0.812 ± 0.006, confirming that semantic enhancement improves diagnostic interpretability. Keywords: cognitive diagnosis; large language model; LLM; graph neural network; blended learning. DOI: 10.1504/IJICT.2026.10082047
Abstract: E-commerce dynamic pricing faces challenges such as the prior unknown demand function and continuously shifting consumer preferences. This paper proposes a reinforcement learning-driven approach for e-commerce dynamic pricing management and consumer behaviour inference. The method employs maximum entropy inverse reinforcement learning to infer the implicit utility function of consumers from historical purchase logs, eliminating reliance on prior knowledge. Next, a proximal policy optimisation-based solver is used, incorporating the inferred expected demand into the reward shaping mechanism to alleviate the cold start problem with sparse rewards. Finally, model-agnostic meta-learning is introduced to enable the strategy to acquire universally initialised parameters across scenarios, requiring only a few gradient updates to quickly adapt to new scenarios. Experimental outcome indicates that the suggested approach achieves a cumulative gross profit margin of 29.7%, which is 3.6% higher than the best baseline, and attains optimal performance in inventory turnover, stockout rate, and price stability. Keywords: e-commerce dynamic pricing management; consumer behaviour inference; reinforcement learning; proximal policy optimisation; model-agnostic meta-learning. DOI: 10.1504/IJICT.2026.10081860
Abstract: Digital generation of ink-wash animation faces a core challenge in balancing style fidelity and temporal coherence. Existing style transfer methods are limited to static images, while video diffusion models lack control over brushwork and ink texture of traditional Chinese painting. We propose a temporal diffusion model with style consistency constraints, comprising a style embedding module extracting ink texture and tonal layers, a temporal consistency module enforcing motion smoothness via inter-frame attention and optical flow, and a bidirectional constrained diffusion module incorporating style and temporal guidance during denoising. Experiments on densely annotated video segmentation and University of Central Florida 101 shows that compared with stable video diffusion, our method reduces Fréchet Video Distance by 13.8% to 276.4, improves style score from 0.62 to 0.89, and maintains content fidelity. This model provides an effective pathway for ink-wash animation, achieving artistic expressiveness and natural motion. Keywords: ink animation generation; diffusion model; style consistency; temporal constraints video generation. DOI: 10.1504/IJICT.2026.10082049 |
Open Access
Keep up-to-date
-
Our Blog -
Join us on Bluesky -
Follow us on X -
Visit us on Facebook - Our Newsletter (subscribe for free)
- RSS Feeds
- New issue alerts
- Inderscience is a member of publishing organisations including:
-
-
