Yunxiang Zheng
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yunxiang Zheng has a total of 71 co-authors. Here are the top co-authors by number of shared publications:
- Jingxiu Huang
School of Educational Information Technology, South China Normal University, Guangzhou 510631, China
14 shared publications Jingxiu Huang
South China Normal University
4 shared publicationsJinping Zhong
School of Educational Information Technology, South China Normal University, Guangzhou, China
4 shared publicationsJunyi Zhang
School of Educational Information Technology, South China Normal University, Guangzhou, China
4 shared publicationsLixiang Liu
School of Educational Information Technology, South China Normal University, Guangzhou, China
3 shared publications- Lixiang Liu
School of Educational Information Technology, South China Normal University, No. 55 Western Zhongshan Avenue, Guangzhou 510631, China
3 shared publications
Latest Publications
- Open Access
- 11 September 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (18) , 9048
Effects of a Digital Game-Based Curriculum on Students’ Digital Rights and Responsibilities Literacy, Cognitive Absorption and Learning Anxiety
Digital rights and responsibilities education is a core component of digital citizenship in primary schools, yet existing approaches often prioritize rule transmission over behavioral internalization. This study examined the effects of integrating Digital Game-Based Learning (DGBL) into the final lesson of a digital rights and responsibilities curriculum. A quasi-experimental study was conducted with 88 fifth-grade students during a four-lesson curriculum delivered over four weeks. Both groups received the same teacher-led instruction during the first three lessons. In the final lesson, Cyber Judge was incorporated into instruction for the experimental group (EG) as a game-supported learning activity, while the control group (CG) received conventional instruction on the same content. Data were analyzed using ANCOVA and t-tests, with Holm–Bonferroni adjustments applied for multiple comparisons. The EG showed higher adjusted post-test digital rights and responsibilities literacy, with significant differences in Rights Awareness and Responsibilities Behavior, and higher overall cognitive absorption, particularly in Temporal Separation. No significant between-group differences were found in learning anxiety. These findings highlight the potential of DGBL to connect ethical knowledge with situated decision-making while underscoring the importance of considering learners’ affective responses in the design of game-supported instruction.
- 11 September 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (18) , 9048
Open AccessEffects of a Digital Game-Based Curriculum on Students’ Digital Rights and Responsibilities Literacy, Cognitive Absorption and Learning Anxiety
Digital rights and responsibilities education is a core component of digital citizenship in primary schools, yet existing approaches often prioritize rule transmission over behavioral internalization. This study examined the effects of integrating Digital Game-Based Learning (DGBL) into the final lesson of a digital rights and responsibilities curriculum. A quasi-experimental study was conducted with 88 fifth-grade students during a four-lesson curriculum delivered over four weeks. Both groups received the same teacher-led instruction during the first three lessons. In the final lesson, Cyber Judge was incorporated into instruction for the experimental group (EG) as a game-supported learning activity, while the control group (CG) received conventional instruction on the same content. Data were analyzed using ANCOVA and t-tests, with Holm–Bonferroni adjustments applied for multiple comparisons. The EG showed higher adjusted post-test digital rights and responsibilities literacy, with significant differences in Rights Awareness and Responsibilities Behavior, and higher overall cognitive absorption, particularly in Temporal Separation. No significant between-group differences were found in learning anxiety. These findings highlight the potential of DGBL to connect ethical knowledge with situated decision-making while underscoring the importance of considering learners’ affective responses in the design of game-supported instruction.
- Open Access
- 13 August 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (16) , 8072
A GCN and Autoencoder-Based Framework for Fine-Grained Diagnosis of Learners’ Cognitive Structures
Accurately diagnosing fine-grained differences in learners’ cognitive structures, which arise from their understanding of interdependencies among knowledge concepts, remains a significant challenge in intelligent education. To address this gap, this paper introduces a novel cognitive diagnosis framework that integrates Graph Convolutional Networks (GCNs) with autoencoders. The framework explicitly models the educational prerequisite relationships within a knowledge concept map to infer the structural features of a learner’s cognition. The concept map is first embedded into a structural matrix using struc2vec. A two-layer GCN then refines this matrix by integrating learner-specific mastery evidence derived from individual performance data using the Attribute Hierarchy Method (AHM), to generate a cognitive structure representation. Subsequently, an autoencoder compresses this representation into a low-dimensional vector, enabling the derivation of a cognitive structure coefficient. Empirical analyses of two real-world course datasets (XML: N = 65; Java: N = 38) showed strong associations between the proposed coefficient and course assessment scores (XML: Pearson’s r = 0.803, p < 0.001; Java: r = 0.776, p < 0.001). Case analyses further showed that learners with identical or similar aggregate scores could exhibit distinct concept-level profiles, providing the granularity needed for personalized learning interventions.
- 13 August 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (16) , 8072
Open AccessA GCN and Autoencoder-Based Framework for Fine-Grained Diagnosis of Learners’ Cognitive Structures
Accurately diagnosing fine-grained differences in learners’ cognitive structures, which arise from their understanding of interdependencies among knowledge concepts, remains a significant challenge in intelligent education. To address this gap, this paper introduces a novel cognitive diagnosis framework that integrates Graph Convolutional Networks (GCNs) with autoencoders. The framework explicitly models the educational prerequisite relationships within a knowledge concept map to infer the structural features of a learner’s cognition. The concept map is first embedded into a structural matrix using struc2vec. A two-layer GCN then refines this matrix by integrating learner-specific mastery evidence derived from individual performance data using the Attribute Hierarchy Method (AHM), to generate a cognitive structure representation. Subsequently, an autoencoder compresses this representation into a low-dimensional vector, enabling the derivation of a cognitive structure coefficient. Empirical analyses of two real-world course datasets (XML: N = 65; Java: N = 38) showed strong associations between the proposed coefficient and course assessment scores (XML: Pearson’s r = 0.803, p < 0.001; Java: r = 0.776, p < 0.001). Case analyses further showed that learners with identical or similar aggregate scores could exhibit distinct concept-level profiles, providing the granularity needed for personalized learning interventions.
- Open Access
- 21 May 2026
- Published by Springer Nature in Research Square
Automatically detecting cognitive levels of teachers’ questions in classroom discourse: Insights from interpretable artificial intelligence
Effective classroom questioning is crucial for enhancing student learning. While existing studies have categorized teachers’ questions, most relied on manual methods lacking interpretability. This study addresses this gap by employing interpretable machine learning to automatically classify teachers’ questions according to Bloom’s taxonomy and to uncover the linguistic features that distinguish different cognitive levels.A total of 5,232 questions from 86 course videos were manually coded and analyzed. We evaluated the performance of decision tree and five tree-based classification models on identifying the cognitive level of teacher questions. SHapley Additive exPlanations was used to discern each feature’s contribution to the model’s predictions. Results showed significant linguistic differences across cognitive levels (e.g., pronouns, adverbs, causal words). Random Forest and LightGBM performed best, with LightGBM selected for interpretability analysis considering both macro and micro performance indexes comprehensively. The study demonstrates interpretable artificial intelligence can effectively classify question types and reveal linguistic patterns. These findings offer a foundation for developing automated feedback tools to help teachers refine their questioning strategies and support professional development.
- 21 May 2026
- Published by Springer Nature in Research Square
Open AccessAutomatically detecting cognitive levels of teachers’ questions in classroom discourse: Insights from interpretable artificial intelligence
Effective classroom questioning is crucial for enhancing student learning. While existing studies have categorized teachers’ questions, most relied on manual methods lacking interpretability. This study addresses this gap by employing interpretable machine learning to automatically classify teachers’ questions according to Bloom’s taxonomy and to uncover the linguistic features that distinguish different cognitive levels.A total of 5,232 questions from 86 course videos were manually coded and analyzed. We evaluated the performance of decision tree and five tree-based classification models on identifying the cognitive level of teacher questions. SHapley Additive exPlanations was used to discern each feature’s contribution to the model’s predictions. Results showed significant linguistic differences across cognitive levels (e.g., pronouns, adverbs, causal words). Random Forest and LightGBM performed best, with LightGBM selected for interpretability analysis considering both macro and micro performance indexes comprehensively. The study demonstrates interpretable artificial intelligence can effectively classify question types and reveal linguistic patterns. These findings offer a foundation for developing automated feedback tools to help teachers refine their questioning strategies and support professional development.
- Open Access
- 5 April 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (7) , 3552
Design and Evaluation of a Question-Answering System Based on Knowledge Graph-Augmented Large Language Models in K–12 Artificial Intelligence Curriculum
Digital transformation is reshaping the education sector, fostering an AI-enabled, learner-centered ecosystem. This shift is characterized by the adoption of large language models (LLMs) in education, which is forging a new paradigm for intelligent teaching. However, the integration of LLMs into K–12 AI education is often hindered by their tendency to generate factually inaccurate and pedagogically misaligned content. To address this, we constructed a knowledge graph (KG) of the K–12 AI curriculum and developed a question-answering system based on KG-augmented LLMs. The system was evaluated on a dedicated AI curriculum dataset comprising 1098 questions categorized into three difficulty levels. The evaluation employed the G-Eval with no-reference metrics. Using DeepSeek-V3 as the scoring model, the system performance was assessed across three mainstream LLMs and measured along five distinct dimensions. Results indicated that the integration of curriculum KG significantly enhanced the factual accuracy and relevance of LLM-generated answers in K–12 AI education. However, this enhancement involves a trade-off, as the incorporation of non-declarative knowledge can negatively affect linguistic fluency and coherence. Performance gains varied across LLMs: Qwen and Baichuan demonstrated the strongest improvements, particularly in complex tasks. This study provides a scalable, knowledge-anchored framework for developing reliable AI teaching assistants, demonstrating a practical pathway to mitigate domain-specific hallucinations in educational applications.
- 5 April 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (7) , 3552
Open AccessDesign and Evaluation of a Question-Answering System Based on Knowledge Graph-Augmented Large Language Models in K–12 Artificial Intelligence Curriculum
Digital transformation is reshaping the education sector, fostering an AI-enabled, learner-centered ecosystem. This shift is characterized by the adoption of large language models (LLMs) in education, which is forging a new paradigm for intelligent teaching. However, the integration of LLMs into K–12 AI education is often hindered by their tendency to generate factually inaccurate and pedagogically misaligned content. To address this, we constructed a knowledge graph (KG) of the K–12 AI curriculum and developed a question-answering system based on KG-augmented LLMs. The system was evaluated on a dedicated AI curriculum dataset comprising 1098 questions categorized into three difficulty levels. The evaluation employed the G-Eval with no-reference metrics. Using DeepSeek-V3 as the scoring model, the system performance was assessed across three mainstream LLMs and measured along five distinct dimensions. Results indicated that the integration of curriculum KG significantly enhanced the factual accuracy and relevance of LLM-generated answers in K–12 AI education. However, this enhancement involves a trade-off, as the incorporation of non-declarative knowledge can negatively affect linguistic fluency and coherence. Performance gains varied across LLMs: Qwen and Baichuan demonstrated the strongest improvements, particularly in complex tasks. This study provides a scalable, knowledge-anchored framework for developing reliable AI teaching assistants, demonstrating a practical pathway to mitigate domain-specific hallucinations in educational applications.
- Open Access
- 30 September 2025
- Published by Springer Nature in Research Square
Machines flunking an exam: Evaluating large language models on course-related open questions
Large language models (LLMs) have grown rapidly in China since the rise of ChatGPT, touching many fields. These models offer a promising solution for responding to students’ questions while learning. However, most relevant research has addressed English-language and multiple-choice questions, namely on general or medical topics; LLMs’ performance in languages such as Chinese and in answering course-related open questions is less clear. To evaluate the performance of LLMs for answering course-related open questions in Chinese, this study explores how well LLMs respond to open-ended, course-specific queries in Chinese (glossary, short-answer, and essay questions). Answers from six LLMs were evaluated based on expert- and machine-assigned grades. Correlation analysis revealed which metrics were most efficacious in automated assessment. We next compared each LLM’s scores to determine the three best-performing models, which were then used for comparison with students. Overall, the selected LLMs’ performance was unsatisfactory: these models demonstrated lower scores relative to students, especially on glossary questions. We have thus recommended several ways to refine LLMs’ operation. These suggestions can help to promote the spread of LLMs and serve as a reference for students and educators when asking course-related open questions. The implications, limitations, and research directions arising from our study are also discussed.
- 30 September 2025
- Published by Springer Nature in Research Square
Open AccessMachines flunking an exam: Evaluating large language models on course-related open questions
Large language models (LLMs) have grown rapidly in China since the rise of ChatGPT, touching many fields. These models offer a promising solution for responding to students’ questions while learning. However, most relevant research has addressed English-language and multiple-choice questions, namely on general or medical topics; LLMs’ performance in languages such as Chinese and in answering course-related open questions is less clear. To evaluate the performance of LLMs for answering course-related open questions in Chinese, this study explores how well LLMs respond to open-ended, course-specific queries in Chinese (glossary, short-answer, and essay questions). Answers from six LLMs were evaluated based on expert- and machine-assigned grades. Correlation analysis revealed which metrics were most efficacious in automated assessment. We next compared each LLM’s scores to determine the three best-performing models, which were then used for comparison with students. Overall, the selected LLMs’ performance was unsatisfactory: these models demonstrated lower scores relative to students, especially on glossary questions. We have thus recommended several ways to refine LLMs’ operation. These suggestions can help to promote the spread of LLMs and serve as a reference for students and educators when asking course-related open questions. The implications, limitations, and research directions arising from our study are also discussed.
- Open Access
- 8 August 2025
- Published by Springer Nature in Humanities and Social Sciences Communications
- Vol. 12 (1) , 1-13
Beyond anger: uncovering complex emotional patterns between cyberbullying roles through affective computing and epistemic network analysis
Although emotions are regarded as essential in automatic cyberbullying detection, the nuanced links between emotion types and roles remain underexplored. The dynamics of cyberbullying are therefore somewhat ambiguous. To address these issues, we analyzed the emotional patterns and connections between five cyberbullying roles (bullies, outsiders, assistants, defenders, and reporters) on a Chinese social media platform. Six emotions were extracted from 11,601 comments using a large pre-trained model for affective computing. Through epistemic network analysis, this study identified three co-occurrence patterns of emotional expressions among these roles, namely, anger-dominated negative pattern, happiness-anger conflicting pattern, and surprise-fear moderate pattern. Beyond just Angry, three emotions (Fearful, Happy, and Surprised) varied significantly among nearly all roles. In addition to the valence of emotions, the position of these roles within the overall network may also be associated with different levels of emotional arousal. Results of subtracted networks for three role pairs further indicated that these emotional co-occurrences may help identify roles for their perceptions, judgments, and intentions regarding others. These insights hold promise for enhancing targeted bullying detection and intervention.
- 8 August 2025
- Published by Springer Nature in Humanities and Social Sciences Communications
- Vol. 12 (1) , 1-13
Open AccessBeyond anger: uncovering complex emotional patterns between cyberbullying roles through affective computing and epistemic network analysis
Although emotions are regarded as essential in automatic cyberbullying detection, the nuanced links between emotion types and roles remain underexplored. The dynamics of cyberbullying are therefore somewhat ambiguous. To address these issues, we analyzed the emotional patterns and connections between five cyberbullying roles (bullies, outsiders, assistants, defenders, and reporters) on a Chinese social media platform. Six emotions were extracted from 11,601 comments using a large pre-trained model for affective computing. Through epistemic network analysis, this study identified three co-occurrence patterns of emotional expressions among these roles, namely, anger-dominated negative pattern, happiness-anger conflicting pattern, and surprise-fear moderate pattern. Beyond just Angry, three emotions (Fearful, Happy, and Surprised) varied significantly among nearly all roles. In addition to the valence of emotions, the position of these roles within the overall network may also be associated with different levels of emotional arousal. Results of subtracted networks for three role pairs further indicated that these emotional co-occurrences may help identify roles for their perceptions, judgments, and intentions regarding others. These insights hold promise for enhancing targeted bullying detection and intervention.
- 11 June 2025
- Published by Taylor & Francis in Interactive Learning Environments
- Vol. 34 (3) , 1062-1081
How the timing of using digital games in courses influences students' learning: a study on digital citizenship education
- 11 June 2025
- Published by Taylor & Francis in Interactive Learning Environments
- Vol. 34 (3) , 1062-1081
How the timing of using digital games in courses influences students' learning: a study on digital citizenship education
- Open Access
- 26 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 94169-94185
A Cognitive Topic Modeling Approach for Detailed Learner Analysis in Asynchronous Discussions
Asynchronous discussion platforms are crucial components of online education: these sites promote knowledge exchange and enhance the learning experience. Students’ discussion data contain abundant information, including personal interests and preferences. Such content also offers insight into learners’ cognitive states. Uncovering the features of various cognitive processes holds value for the education field. We thus propose a cognitive topic modeling method for asynchronous online discussions called SLPoK–ENA. This approach consists of two parts: 1) SLPoK, a topic modeling technique that uses sentence embedding to improve latent Dirichlet allocation (LDA); and 2) the integration of SLPoK and epistemic network analysis (ENA) to perform cognitive topic modeling. Results from two real-world course datasets show that SLPoK outperforms LDA in topic modelling when dealing with educational data, achieving Normalized Mutual Information (NMI) values of 0.62 (Dataset 1) and 0.75 (Dataset 2), Adjusted Rand Index (ARI) values of 0.45 and 0.73, Silhouette Coefficient (SC) of 0.225 and 0.308, and Calinski-Harabasz Index (CH) of 196.350 and 464.415 on the two datasets, respectively —all significantly higher than the traditional LDA baseline model. Also, this method is better at capturing students’ cognitive characteristics. Experiments demonstrate that SLPoK–ENA can identify learners’ topic preferences under different cognitive conditions. This method has practical value for understanding students and improving teaching quality in online education settings.
- 26 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 94169-94185
Open AccessA Cognitive Topic Modeling Approach for Detailed Learner Analysis in Asynchronous Discussions
Asynchronous discussion platforms are crucial components of online education: these sites promote knowledge exchange and enhance the learning experience. Students’ discussion data contain abundant information, including personal interests and preferences. Such content also offers insight into learners’ cognitive states. Uncovering the features of various cognitive processes holds value for the education field. We thus propose a cognitive topic modeling method for asynchronous online discussions called SLPoK–ENA. This approach consists of two parts: 1) SLPoK, a topic modeling technique that uses sentence embedding to improve latent Dirichlet allocation (LDA); and 2) the integration of SLPoK and epistemic network analysis (ENA) to perform cognitive topic modeling. Results from two real-world course datasets show that SLPoK outperforms LDA in topic modelling when dealing with educational data, achieving Normalized Mutual Information (NMI) values of 0.62 (Dataset 1) and 0.75 (Dataset 2), Adjusted Rand Index (ARI) values of 0.45 and 0.73, Silhouette Coefficient (SC) of 0.225 and 0.308, and Calinski-Harabasz Index (CH) of 196.350 and 464.415 on the two datasets, respectively —all significantly higher than the traditional LDA baseline model. Also, this method is better at capturing students’ cognitive characteristics. Experiments demonstrate that SLPoK–ENA can identify learners’ topic preferences under different cognitive conditions. This method has practical value for understanding students and improving teaching quality in online education settings.
- Open Access
Does Part of Speech Have an Influence on Cyberbullying Detection?
With the development of the Internet, the issue of cyberbullying on social media has gained significant attention. Cyberbullying is often expressed in text. Methods of identifying such text via machine learning have been growing, most of which rely on the extraction of part-of-speech (POS) tags to improve their performance. However, the current study only arbitrarily used part-of-speech labels that it considered reasonable, without investigating whether the chosen part-of-speech labels can better enhance the effectiveness of the cyberbullying detection task. In other words, the effectiveness of different part-of-speech labels in the automatic cyberbullying detection task was not proven. This study aimed to investigate the part of speech in statements related to cyberbullying and explore how three classification models (random forest, naïve Bayes, and support vector machine) are sensitive to parts of speech in detecting cyberbullying. We also examined which part-of-speech combinations are most appropriate for the models mentioned above. The results of our experiments showed that the predictive performance of different models differs when using different part-of-speech tags as inputs. Random forest showed the best predictive performance, and naive Bayes and support vector machine followed, respectively. Meanwhile, across the different models, the sensitivity to different part-of-speech tags was consistent, with greater sensitivity shown towards nouns, verbs, and measure words, and lower sensitivity shown towards adjectives and pronouns. We also found that the combination of different parts of speech as inputs had an influence on the predictive performance of the models. This study will help researchers to determine which combination of part-of-speech categories is appropriate to improve the accuracy of cyberbullying detection.
Open AccessDoes Part of Speech Have an Influence on Cyberbullying Detection?
With the development of the Internet, the issue of cyberbullying on social media has gained significant attention. Cyberbullying is often expressed in text. Methods of identifying such text via machine learning have been growing, most of which rely on the extraction of part-of-speech (POS) tags to improve their performance. However, the current study only arbitrarily used part-of-speech labels that it considered reasonable, without investigating whether the chosen part-of-speech labels can better enhance the effectiveness of the cyberbullying detection task. In other words, the effectiveness of different part-of-speech labels in the automatic cyberbullying detection task was not proven. This study aimed to investigate the part of speech in statements related to cyberbullying and explore how three classification models (random forest, naïve Bayes, and support vector machine) are sensitive to parts of speech in detecting cyberbullying. We also examined which part-of-speech combinations are most appropriate for the models mentioned above. The results of our experiments showed that the predictive performance of different models differs when using different part-of-speech tags as inputs. Random forest showed the best predictive performance, and naive Bayes and support vector machine followed, respectively. Meanwhile, across the different models, the sensitivity to different part-of-speech tags was consistent, with greater sensitivity shown towards nouns, verbs, and measure words, and lower sensitivity shown towards adjectives and pronouns. We also found that the combination of different parts of speech as inputs had an influence on the predictive performance of the models. This study will help researchers to determine which combination of part-of-speech categories is appropriate to improve the accuracy of cyberbullying detection.
- Top Cited PapersOpen Access
Effects of digital game-based learning on students’ digital etiquette literacy, learning motivations, and engagement
Top Cited PapersOpen AccessEffects of digital game-based learning on students’ digital etiquette literacy, learning motivations, and engagement