Traditional methods for generating course content rely on manual rules, lacking dynamic knowledge support and semantic monitoring, leading to deviations between generated content and teaching objectives. Therefore, this study proposes an AI-powered framework for generating art and design course content that integrates knowledge graphs and multimodal semantic information extraction. Results show that, in tests on a self-built dataset, the framework achieved knowledge accuracy of 91.21%, knowledge graph relationship consistency of 89.91%, and graph-text relevance of 0.33%, representing average improvements of 39.3%, 29.8%, and 13.8% compared to the best baseline. The model converges quickly with a response time as low as 1.0 millisecond, a service error rate of only 4%, and maintains 92% stability under 100 concurrent requests. This framework demonstrates excellent scalability and service robustness, providing core technical support for the transformation of art and design education in the era of artificial intelligence.
Automatic scoring of English essays is important for teaching evaluation and large-scale assessment, but existing methods often insufficiently capture discourse logic and multi-dimensional writing features. This study proposes an automatic scoring model based on topic granularity segmentation and feature extraction. First, a CNN-BiLSTM-CRF framework is constructed to segment essays into coherent topic units and capture topic development logic. Then, RoBERTa is used to extract semantic, lexical, syntactic, and structural features, which are integrated by LightGBM for final scoring. Experimental results show that the proposed model achieves a correlation coefficient of 0.9613 with human scoring, maintains scoring accuracy above 95.1% across different essay types, obtains an F1 score of 95.6%, precision of 94.8%, and a mean absolute error of 0.67. Its maximum memory usage is 394 MB, and the response time for 2,000 essays is 6.24 s, demonstrating high scoring accuracy and efficiency.
With the increasing demand for personalised learning support and intelligent recommendation in engineering education, traditional learning path planning methods still face challenges in handling multi-source learner profiles and complex course associations, resulting in low path matching rates and high recommendation latency. To address these issues, this study proposes a learning path recommendation framework integrating knowledge graph modelling and multi-source behaviour perception. The framework combines dynamic knowledge tracing with hierarchical regularised user modelling and knowledge graph embedding with an upper confidence bound strategy to achieve accurate path recommendation. Experimental results show that the proposed model achieves an F1-score of 93.22% on a public dataset, while the average response delay is within 2 seconds. Under noise disturbances, the fluctuation of path similarity remains within +/- 2.5%, and the course completion rate reaches 97.3% across four engineering disciplines, demonstrating strong accuracy, real-time performance and adaptability.
To address the challenge of monitoring student engagement in mechanical engineering classrooms, this study proposes an enhanced multi-task cascaded convolutional neural network combined with particle filtering and a support vector machine. The model integrates facial expression and head pose recognition to evaluate classroom attention. Experimental results show face detection accuracy reached 91% on the FDDB dataset, while expression recognition achieved 92% accuracy and 90% F1-score on CK+. The combined method improved behaviour recognition accuracy by 13.28% and increased frame rates by 13.6-25% compared to single-method approaches. Practical application demonstrated a 22% increase in homework completion, 28.46% rise in assignment scores, and doubled interaction frequency. The model effectively assesses teaching effectiveness and provides data-driven support for improving instructional content and student engagement.
This paper proposes a segmentation algorithm integrating a spatial attention mechanism with a mask R-CNN to address safety risks in dynamic physical education scenes and improve motion target segmentation. The method first performs target segmentation using mask R-CNN, enhances feature representation through spatial attention, and completes monitoring and localisation via a visual processor and a conditional convolution instance segmentation model. Experiments show strong performance: a segmentation boundary value of 0.967, F1-score of 96.75%, and 3.27% average absolute error on the YouTube-VOS dataset, with false negative and false positive rates of 1.47% and 1.81%. On the CDnet 2014 dataset, pixel accuracy reaches 95.42% with an iteration time of 4.95 s. On a self-constructed dataset, the method achieves 0.972 average precision and 0.984 panoptic quality. These results demonstrate accurate and efficient motion target segmentation, supporting intelligent physical education applications.
This paper proposes an immersive learning system based on mixed reality technology to address the lack of authenticity, interaction, and personalisation in traditional business English teaching. The system integrates 3D scene modelling, multimodal interaction, and reinforcement learning-driven adaptive path planning. Experimental results show stable technical performance with 58.2 FPS, 93.4% gesture recognition accuracy, and 97.8% conflict resolution success rate. Teaching effectiveness includes a 42.7% improvement in oral fluency, a 146.7% increase in terminology use, and 82.4% long-term memory retention. The system enhances learners' language application and practical skills.
Driven by big data and artificial intelligence, online teaching tools like virtual labs are emerging, yet they often suffer from low resource utilisation, limited load capacity, and high costs. To overcome these challenges, this study develops a cloud-based infrastructure leveraging the on-demand and elastic features of cloud computing. An enhanced load-balancing algorithm integrating particle swarm optimisation and cuckoo search is proposed, along with a virtual-machine migration strategy, to build a virtual teaching lab. Results show that the algorithm achieves 12.35% CPU utilisation, 17.23% memory utilisation, 33.02% bandwidth utilisation, and a maximum response delay of 210.11 ms. The migration strategy also proves more cost-effective and efficient than alternatives. In practice, 98% of students improved their experimental skills using the virtual lab, and 52.72% scored above 90, outperforming students in traditional lab teaching. The virtual lab demonstrates stable loading, high efficiency, low cost, and supports educational digital transformation.
In the context of diversified vocational training scenarios, traditional static evaluation is difficult to control teaching quality fluctuations and anomalies in real time and accurately. Therefore, a method has been proposed to integrate the backpropagation neural network model with an improved particle swarm optimisation algorithm. Quantify the impact of input features on teaching quality through MIV, screen key features to reduce data dimensionality, and dynamically adjust the search step size of particle swarm optimisation algorithm accordingly. Capture nonlinear relationships through backpropagation neural networks and improve global optimisation capabilities through particle swarm optimisation. In precision testing, the accuracy of the research model on the test set reached 99.56%. The error rate significantly decreased from the initial 4.02% to the final 2.13%, indicating its strong generalisation ability. This model solves the problem of failing to identify potential teaching risks. It provides a new method for teaching quality management in vocational training.
The current immersive learning environment for English still cannot meet the practical needs of language transfer and intensive use in terms of the authenticity of language scenes, real-time response of learning behaviour, and semantic continuity of contextual interaction. Therefore, this paper proposes an educational digital twin modelling method. This paper first constructs a learner digital twin model in a synchronous manner based on multi-source data such as speech features, eye movement trajectories, and operational behaviours. Secondly, this paper constructs a dynamic language scene model by mining the spatial structure and semantic paths of typical language tasks, and finally uses context modelling and semantic graph linking algorithms to perform context analysis on learners' inputs. The experimental results show that the average response time for speech input recognition is 434 ms, and the average matching rate of interactive context is over 90%; the satisfaction score of users when activating interactive context is 86.7.
This paper develops low-cost intelligent devices and designs personalised learning algorithms to optimise the effectiveness of early childhood education driven by artificial intelligence (AI). Firstly, this paper designs a teacher technical literacy training module and uses a virtual reality simulator to enhance teachers' ability to operate artificial intelligence tools. Secondly, based on children's behavioural data, this paper applies collaborative filtering algorithms and long short-term memory (LSTM) models to construct an adaptive learning system. Finally, with the help of 3D modelling software and spatial audio technology, this paper constructs a virtual reality interactive scene to enhance the immersive learning experience, improves the security and usability of the human-computer interaction interface through natural language processing models and touch interaction optimisation. The research results indicate that in interactive teaching scenarios, the AR rendering delay of high-end devices is only 25 ms, while the AR rendering delay of low-end devices is 40 ms.
Interactive teaching practices improve the understandability of lessons and subjects through audio-visual representations with human-computer interaction (HCI) serving as a foundational enabler. In particular, voice-assisted interfaces facilitate natural, hands-free information exchange between students and digital systems, allowing real-time feedback, note recognition, and adaptive instruction in music classrooms. This teaching is backboned with human-computer interactions for touch and voice-assisted interfaces for information exchange. In this article, an itinerary interaction module for note procedures (IIM-NP) in music classrooms is designed for improving the understandability and applications of music notes. This method first stores voluptuous musical notes for introduction, understanding, and application of tones. The stored notes are filtered based on the student's understandability aiding ease of teaching. In the processing phase, the understandable and hard note teaching practices are differentiated for which itinerary interactions are planned. This planning relies on teaching recommendations as provided by the state learning. The understandable and hard note teaching interactions are transited based on the itinerary steps pursued. In the transition changes, the understandability level serves as the reward factor from which the procedures are simplified or improved for further interactions. The different transitions balance the understandability levels of students of different ages and learning abilities.
Today's society is an information-based and open society. Social informatisation makes school education enter a comprehensive open teaching. An open society requires teachers to face educational reform with a broad vision. Looking at the basic education reform in the world today, the main goal is to cultivate students' overall quality, creativity and practical ability, which is what China's current basic education lacks. Through the research on open education, this paper mainly explained its characteristics, direction and principles. Through the analytic hierarchy process (AHP), the teaching situation was scored, and the classroom atmosphere and students' learning situation of open education and traditional education were compared. The results showed that the open education based on outcome-based education (OBE) education concept had a better classroom atmosphere. Compared with traditional education, the number of students who failed in open education had decreased by 20%, and the number of students with excellent conditions had increased by 40%, which also showed that open education was more suitable for teaching.
To address noise, speech masking, and weak robustness in classroom emotion recognition, this study proposes a model combining enhanced segmentation clustering and multi-feature fusion. An improved U-Net with local loss supervision first performs denoising. Secondly, using MFCC features combined with Bayesian segmentation and K-means clustering to process speech signals. Finally, MFCC, formant, and pitch features are integrated into an attention-based BiLSTM for emotion recognition. Results show the U-Net achieved a loss of 0.26 after 16 iterations, with PESQ at 3.01 and STOI 84.21%. Segmentation false negative and positive rates were 13.84% and 12.52%. K-means purity reached 90.85%. The multi-feature model attained 92.98% accuracy for excited emotion, and the full system reached 93.62% test accuracy. The model improves recognition in complex classrooms, supporting personalised smart education.
