International Journal of Information and Communication Technology (11 papers in press)

Regular Issues

• Open Access

  Identification of translation bias in Chinese-Korean Confucian texts based on pre-trained language models ( Free Full-text Access )

  by Zhengfeng Huang 
  Abstract: Confucian classics hold a foundational position in the history of Sino-Korean cultural exchange. However, machine translation of these texts often leads to semantic distortion and cultural bias. This paper proposes an automated bias identification framework based on the pre-trained cross-lingual model x-language model-robustly optimised bidirectional encoder representations from transformers pretraining approach. Through a multi-task architecture integrates contrastive learning, semantic role labelling, and context-aware alignment, our method effectively identifies and quantifies semantic, cultural, and grammatical deviations in translated Confucian texts. Experimental results on multiple public available corpora demonstrate that the proposed approach achieves an F1-score of 0.83 and accuracy of 85%, outperforming existing baselines in both metrics, especially in identifying culturally specific terms and nuanced expressions (F1 = 0.86 for cultural bias). This research provides valuable methodological insights for evaluating classical text translation quality and supports the accurate dissemination and digital preservation of Confucian cultural heritage.
  Keywords: pre-trained language models; PLMs; Chinese-Korean translation; Confucian texts; bias identification; cross-language processing.
  DOI: 10.1504/IJICT.2025.10074595
   
  

• Open Access

  Real-time detection of business English grammar errors driven by transfer learning ( Free Full-text Access )

  by Zhenxin Fang, Zhenyu Song 
  Abstract: Improving the grammatical accuracy of Business English writing is crucial, but general grammar checking tools often struggle to adapt to professional contexts. This study proposes a real-time grammar error detection method based on BERT transfer learning, aimed at enhancing performance in business scenarios. Methodologically, the BERT-base pre-trained model is directly utilised to capture general language features. To meet real-time requirements, a lightweight model inference architecture was designed. Experimental results show that the model fine-tuned for the business domain achieves an accuracy rate of 89.2% and an F1 score of 0.842. The improvements are particularly significant in detecting formal expressions and complex sentence structures specific to business texts. This study demonstrates that combining BERT-based transfer learning with fine-tuning using small yet representative domain-specific datasets can effectively enhance the practicality and accuracy of grammar error detection in Business English.
  Keywords: transfer learning; business English; grammar error detection; BERT.
  DOI: 10.1504/IJICT.2025.10074621
   
  

• Open Access

  A spatio-temporal transformer predictive model for elderly-oriented tourism via attention mechanism ( Free Full-text Access )

  by Jiya Sun 
  Abstract: To address the issue that current models for predicting the potential of retirement destinations overlook the spatio-temporal correlations between influencing factors, this paper first selects the influencing factors of retirement destination potential and designs an improved empirical mode decomposition algorithm to decompose these factors, obtaining the individual mode components. Then, the characteristics of each mode component are captured, and the spatio-temporal dependencies are unified through an adaptive embedding mechanism. Subsequently, a temporal self-attention module is designed to capture temporal dependencies, and a spatial self-attention mechanism is implemented to model geographical relationships. Feature fusion is achieved using a multi-head attention mechanism, and the prediction results are output through a feedforward neural network. Experimental outcome indicates that the prediction accuracy of the suggested model improves by 2.7%-11.8% compared to the baseline model, validating the superiority of the suggested model.
  Keywords: potential prediction; spatiotemporal transformer; empirical mode decomposition; EMD; attention mechanism.
  
  

• Open Access

  Energy efficiency analysis and optimisation strategies for green building design based on gravitational search algorithm ( Free Full-text Access )

  by Yaxi Gong, Yingyi Ma, Shanshan Cheng 
  Abstract: As peoples requirements for energy saving and emission reduction continue to increase, the issue of energy consumption in buildings has received more and more attention. How to efficiently optimise the energy consumption of green buildings has become an important research goal in the field of energy consumption analysis and architectural design. This study, aiming at the energy consumption problem in green buildings, designs a method based on gravitational search algorithm (GSA) to optimise energy consumption. First, sensor data of equipment in the building is collected. Then, a multi-objective optimisation model is constructed to ensure that the final goal is the lowest energy consumption without reducing comfort. The final experimental results show that the overall building energy use decreased by 33.8% because the GSA algorithm can effectively reduce the overall energy consumption of building equipment and meets the requirements for energy consumption optimisation in green buildings.
  Keywords: green buildings; gravitational search algorithm; GSA; energy consumption analysis; multi-objective optimisation model.
  
  

• Open Access

  Intelligent recognition and analysis system of students' behaviour in continuing education based on classroom video ( Free Full-text Access )

  by Ye Zhiqun 
  Abstract: Student behaviour recognition is crucial for intelligent education but faces challenges in accuracy, robustness under complex conditions like occlusion and lighting variations, and cross-scenario generalisation. This paper proposes the EAST-GCN-HRNet model, which integrates spatiotemporal features and multimodal data to enhance recognition precision and robustness. The model combines HRNet's high-resolution feature extraction, GCN's temporal joint graph modelling, and the EAST module's feature fusion within an end-to-end, multi-scale framework. Experimental results demonstrate the system achieves 86.5% mAP on the SCB-Dataset3 test set, outperforming HRNet by 3.8%. It also shows strong generalisation, with a PCK@0.2 of 63.8% on the AP-10K animal pose dataset (11.5% higher than Hourglass), and robustness with only 4.8% mAP decay in dynamic occlusion scenarios - half that of baseline models. With a real-time inference speed of 28 FPS and a teacher experience rating of 4.6/5, the model provides a reliable tool for intelligent education.
  Keywords: classroom video; students; behaviour recognition; skeleton model.
  DOI: 10.1504/IJICT.2025.10074499
   
  

• Open Access

  Multi-modal similarity feature exchange and structural perception for person re-identification ( Free Full-text Access )

  by Xuefeng Lei 
  Abstract: Visible-infrared person re-identification is crucial for surveillance, aiming to match person images across visible and infrared modalities. However, spectral and style gaps hinder local structure modelling and cross-modal feature alignment. We propose the cross-modality similarity exchange transformer (CSET) to improve both aspects. CSET uses two modality-specific transformer encoders to extract features independently. A similarity exchange mechanism computes intra-modality similarity and cross-modality Jaccard distance, selectively exchanging correlated token features for local alignment and feature complementation. To enhance structural perception, we introduce a multi-relational heterogeneous graph attention mechanism, building a graph from transformer outputs where positional embedding differences define relation levels. Feature aggregation is guided by relational strength to capture fine-grained structural cues. Experiments on RegDB and SYSU-MM01 show CSET outperforms state-of-the-art methods in Rank-1 accuracy and mAP, validating its cross-modal learning effectiveness.
  Keywords: cross-modality person re-identification; similarity exchange; SE; transformer; heterogeneous graph attention; feature alignment.
  DOI: 10.1504/IJICT.2025.10074510
   
  

• Open Access

  Application of distributed artificial intelligence technology in key frame extraction of film and television video ( Free Full-text Access )

  by Feng Cheng 
  Abstract: Traditional video analysis relies on video frames, which often contain redundant data, making key frame extraction essential. However, existing methods frequently suffer from missing or redundant frames. To address this, this paper proposes a video key frame extraction method based on distributed artificial intelligence. First, mutual information between video frames is calculated. Then, SIFT feature points are extracted and transformed into polar coordinates, with each frame divided into sector regions to count feature points and compute inter-frame distances. To enhance precision, the CaffeNet model is adopted as a deep neural network to extract deep features using three training techniques. This approach significantly improves the accuracy of key frame extraction. Experimental results show that the proposed method achieves higher fidelity and compression rates than traditional techniques, and the extracted key frames align closely with reference standards without frame omission, demonstrating its effectiveness and robustness in real-world applications.
  Keywords: distributed artificial intelligence technology; film and television video; key frame extraction; SIFT feature points.
  DOI: 10.1504/IJICT.2025.10074511
   
  

• Open Access

  Intelligent Q&A model construction supported by natural language processing and knowledge graphs ( Free Full-text Access )

  by Xiaoxia Yang 
  Abstract: This paper proposes an intelligent Q&A model that integrates natural language processing and knowledge graph technology. Aiming at the problem of insufficient depth of semantic understanding and weak knowledge relevance of traditional Q&A system, we adopt BERT-based semantic parsing model to realise intent recognition and entity extraction of user questions, and combine with Neo4j graph database to construct multi-source knowledge graph to realise structured knowledge storage; we realise dynamic matching between question vectors and knowledge subgraphs through graph neural network (GNN), and we design multi-jump inference mechanism to improve the ability of answering complex questions. Experiments on the open-domain Q&A dataset show that the model has an accuracy of 90.2% and a recall of 88.7%. The model is validated in educational counselling and medical Q&A scenarios, providing technical support for intelligent services in knowledge-intensive domains.
  Keywords: natural language processing; NLP; knowledge graph; intelligent question answering models; graph neural networks; GNNs.
  DOI: 10.1504/IJICT.2025.10074512
   
  

• Open Access

  Piano performance beat assessment: integrating transformer with multimodal feature learning ( Free Full-text Access )

  by Jun Deng 
  Abstract: This paper proposes PianoTrans-Fusion, a piano performance beat assessment system that integrates the transformer architecture with multimodal feature learning. The system uses three modalities, including audio, video, and MIDI, to perform feature extraction and preprocessing, respectively, and captures fine-grained temporal dependencies in the performance rhythm through multimodal fusion strategies and transformer-based processing modules. Comparative experiments on the MAESTRO dataset show that PianoTrans-Fusion improves rhythm consistency to 0.032 and reduces beat error to 0.071 compared to five baseline methods. Ablation experiments further verify the key roles of transformer, multimodal fusion, and self-attention mechanisms. The results indicate that the system has advantages in terms of accuracy and robustness in beat evaluation, and has application value in intelligent piano accompaniment, music education, and automated performance feedback.
  Keywords: transformer; multimodal feature learning; piano performance; beat assessment.
  DOI: 10.1504/IJICT.2025.10074514
   
  

• Open Access

  Flood disaster prediction using multi-scale deep learning and neuro-fuzzy inference ( Free Full-text Access )

  by Haonan Zhao, Tingjing Xia 
  Abstract: Flood disaster prediction is crucial for disaster prevention and mitigation, but traditional models face dual challenges: insufficient feature extraction and difficulty quantifying uncertainty. This paper proposes multi-scale adaptive neuro-fuzzy inference system. It integrates a multi-scale convolutional feature pyramid network for hierarchical spatiotemporal feature extraction from remote sensing hydrological data with an adaptive neural fuzzy inference system handling rainfall-runoff nonlinear uncertainties. Using global flood alert system and tropical rainfall measuring mission data, experiments in five major river basins (Yangtze, Mississippi, etc.), selected to represent diverse climatic zones and hydrological regimes, show significantly improved 72-hour prediction accuracy, achieving 15-22% root mean square error reduction. Constructed confidence intervals cover 92% of extreme flood events. multi-scale adaptive neuro-fuzzy inference system provides a robust, interpretable tool for smart water management, integrable into real-time flood warning platforms.
  Keywords: flood disaster prediction; multi-scale feature fusion; neural fuzzy inference; spatio-temporal deep learning; uncertainty quantification.
  DOI: 10.1504/IJICT.2025.10074509
   
  

• Open Access

  Intelligent fault diagnosis system for railway infrastructure based on deep learning ( Free Full-text Access )

  by Qi An 
  Abstract: The operational status of railway infrastructure determines the safety of train passage. However, traditional research suffers from low efficiency and difficulty in addressing fault states under variable operating conditions. To address this, this paper first proposes a data balancing method based on improved synthetic minority over-sampling technique and generative adversarial network (GAN) to tackle the imbalance in railway infrastructure signal data. The introduction of unsupervised clustering algorithms and natural neighbour concepts enhances sample generation efficiency. Adding category label information and optimising the training loss function improves the stability of network training. Building upon this foundation, a multi-scale residual network (ResNet) is constructed for feature extraction, mitigating the impact of operational variations on diagnostic outcomes. A subdomain-adaptive transfer learning strategy is employed to achieve fault diagnosis. Experimental validation demonstrates that the proposed method achieves a diagnostic accuracy of 93.86%, delivering highly precise diagnostic results.
  Keywords: railway infrastructure; fault diagnosis; synthetic minority over-sampling technique; generative adversarial network; GAN; transfer learning.
  DOI: 10.1504/IJICT.2025.10074513