Jianwen Sun
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Jianwen Sun has a total of 142 co-authors. Here are the top co-authors by number of shared publications:
- Xiaoxuan Shen
Faculty of Artificial Intelligence in Education, Central China Normal University, Wuhan, China
24 shared publications - Sannyuya Liu
Central China Normal University
16 shared publications - Qing Li
Central China Normal University, Wuhan, China
12 shared publications - Zhi Liu
Faculty of Artificial Intelligence in Education, Central China Normal University, Wuhan, China
8 shared publications - Qian Wan
Central China Normal University, Wuhan, China
6 shared publications - Zhifeng Wang
Department of Digital Media Technology, Central China Normal University, Wuhan, China
6 shared publications
Latest Publications
- 13 October 2026
- Published by Springer Nature
- p. 1-19
Development Context, Research Status, and Future Trends of Learning Sciences in the Chinese Mainland
- 13 October 2026
- Published by Springer Nature
- p. 1-19
Development Context, Research Status, and Future Trends of Learning Sciences in the Chinese Mainland
- Open Access
- 27 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 222-231
FASC: A Feature Aspect-Level Sentiment Consistency Framework for Explainable Recommendation Evaluation
Recent research in explainable recommendation commonly uses natural language explanations to improve transparency and user trust. However, reliably evaluating whether explanations are semantically faithful to users’ multi-dimensional preferences remains challenging. Existing methods mainly rely on text similarity metrics (e.g., BLEU, ROUGE) or shallow feature matching, which are insufficient for assessing whether explanations accurately reflect user preferences, especially in multi-aspect settings. To address these limitations, we propose Feature Aspect-Level Sentiment Consistency (FASC), a framework that quantifies semantic consistency between generated explanations and user-authored reference explanations through aspect coverage and sentiment polarity. FASC uses LLMs as auxiliary tools to extract structured aspect–sentiment units from explanation texts, enabling reproducible metrics for aspect coverage, correctness, and sentiment alignment. To validate the framework, we re-annotated several widely used explainable recommendation datasets to construct benchmarks with fine-grained, aspect-level sentiment labels. We further conducted human studies using pairwise comparisons, showing that FASC aligns more closely with human judgments than traditional metrics. Experimental results indicate that FASC can distinguish subtle differences in semantic faithfulness across models. To support future research, we release all annotated datasets and human evaluation results at https://github.com/YuChenfu1022/FASC.
- 27 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 222-231
Open AccessFASC: A Feature Aspect-Level Sentiment Consistency Framework for Explainable Recommendation Evaluation
Recent research in explainable recommendation commonly uses natural language explanations to improve transparency and user trust. However, reliably evaluating whether explanations are semantically faithful to users’ multi-dimensional preferences remains challenging. Existing methods mainly rely on text similarity metrics (e.g., BLEU, ROUGE) or shallow feature matching, which are insufficient for assessing whether explanations accurately reflect user preferences, especially in multi-aspect settings. To address these limitations, we propose Feature Aspect-Level Sentiment Consistency (FASC), a framework that quantifies semantic consistency between generated explanations and user-authored reference explanations through aspect coverage and sentiment polarity. FASC uses LLMs as auxiliary tools to extract structured aspect–sentiment units from explanation texts, enabling reproducible metrics for aspect coverage, correctness, and sentiment alignment. To validate the framework, we re-annotated several widely used explainable recommendation datasets to construct benchmarks with fine-grained, aspect-level sentiment labels. We further conducted human studies using pairwise comparisons, showing that FASC aligns more closely with human judgments than traditional metrics. Experimental results indicate that FASC can distinguish subtle differences in semantic faithfulness across models. To support future research, we release all annotated datasets and human evaluation results at https://github.com/YuChenfu1022/FASC.
- 16 September 2026
- Published by Elsevier in Neurocomputing
- Vol. 707
Affect-aware modeling of cognitive behavior for explainable knowledge tracing
- 16 September 2026
- Published by Elsevier in Neurocomputing
- Vol. 707
Affect-aware modeling of cognitive behavior for explainable knowledge tracing
- 17 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. PP (4) , 1-14
Leveraging Contrastive Vision-Language Pretraining for Zero-Shot Open Learning Behavior Detection in Smart Education
With the rapid advancement of industrial informatics, educational behavior detection has become crucial for enabling intelligent, personalized teaching in smart education. Traditional methods, reliant on predefined behavior categories, struggle to meet the demands of open-vocabulary learning behavior detection in real-world scenarios. To address this, we introduce education open behavior detection (Edu-OBD), an open-vocabulary learning behavior detection model designed for smart education using a pretraining approach based on region-text pairs. The model employs a contrastive language-image pretraining encoder and contrastive loss to enable effective open-vocabulary behavior detection. To enhance the integration of image and text features, we propose a novel visual-language integrator comprising two modules: text-guided sparse attention (TGSA) and multiscale text enhancer (MSTE). TGSA strengthens category-relevant semantic information, while MSTE refines multiscale feature embeddings. In addition, the position-aware context fusion module (PCFM) uses relative position encoding to improve feature extraction. For evaluation, we developed the zero-shot behavior evaluation dataset (ZBED), tailored for zero-shot evaluation on novel behaviors. Experimental results show that Edu-OBD achieves 33.6% average precision (AP) at 81.2 frames per second (FPS) on ZBED and 28.1% AP for novel category detection, demonstrating its effectiveness as a flexible and efficient solution for educational behavior detection in complex environments, paving the way for future research in smart education.
- 17 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. PP (4) , 1-14
Leveraging Contrastive Vision-Language Pretraining for Zero-Shot Open Learning Behavior Detection in Smart Education
With the rapid advancement of industrial informatics, educational behavior detection has become crucial for enabling intelligent, personalized teaching in smart education. Traditional methods, reliant on predefined behavior categories, struggle to meet the demands of open-vocabulary learning behavior detection in real-world scenarios. To address this, we introduce education open behavior detection (Edu-OBD), an open-vocabulary learning behavior detection model designed for smart education using a pretraining approach based on region-text pairs. The model employs a contrastive language-image pretraining encoder and contrastive loss to enable effective open-vocabulary behavior detection. To enhance the integration of image and text features, we propose a novel visual-language integrator comprising two modules: text-guided sparse attention (TGSA) and multiscale text enhancer (MSTE). TGSA strengthens category-relevant semantic information, while MSTE refines multiscale feature embeddings. In addition, the position-aware context fusion module (PCFM) uses relative position encoding to improve feature extraction. For evaluation, we developed the zero-shot behavior evaluation dataset (ZBED), tailored for zero-shot evaluation on novel behaviors. Experimental results show that Edu-OBD achieves 33.6% average precision (AP) at 81.2 frames per second (FPS) on ZBED and 28.1% AP for novel category detection, demonstrating its effectiveness as a flexible and efficient solution for educational behavior detection in complex environments, paving the way for future research in smart education.
- 18 February 2026
- Published by Elsevier in Applied Soft Computing
- Vol. 193
Leveraging hierarchical attention and dynamic fusion mechanisms for multi-modal speech emotion recognition
- 18 February 2026
- Published by Elsevier in Applied Soft Computing
- Vol. 193
Leveraging hierarchical attention and dynamic fusion mechanisms for multi-modal speech emotion recognition
- Open Access
Mitigating Overthinking in Large Reasoning Models via Difficulty-aware Reinforcement Learning
Large Reasoning Models (LRMs) achieve explicit chain-of-thought expansion by imitating deep thinking behaviors of humans, demonstrating excellent performance in complex task scenarios. However, the deep-thinking mode often leads to unnecessarily lengthy reasoning and resource inefficiency when handling simple tasks. This overthinking phenomenon may arise from the generation preference triggered by the reward function during post-training. Existing research attempts to mitigate overthinking from the perspective of prompt design or model training, but generally underestimates the importance of task difficulty awareness, which makes it difficult for LRMs to effectively allocate reasoning resources. In this paper, we propose Difficulty-aware Policy Optimization (DiPO), a reinforcement learning-based LRM training framework. DiPO encourages LRM to spontaneously model task complexity, and integrates them into reinforcement learning framework to adjust the generation preferences introduced by post-training. A difficulty modeling method based on model self-reasoning is proposed, which significantly reduces the dependence on manual annotation and formalize task complexity. We further develop a difficulty-signal-enhanced reward function that incorporates a penalty for lengthy reasoning while considering reasoning performance and output format. Experimental results indicate that DiPO enables the model to spontaneously adjust inference overhead, significantly reducing redundant tokens without losing performance due to thought compression.
Open AccessMitigating Overthinking in Large Reasoning Models via Difficulty-aware Reinforcement Learning
Large Reasoning Models (LRMs) achieve explicit chain-of-thought expansion by imitating deep thinking behaviors of humans, demonstrating excellent performance in complex task scenarios. However, the deep-thinking mode often leads to unnecessarily lengthy reasoning and resource inefficiency when handling simple tasks. This overthinking phenomenon may arise from the generation preference triggered by the reward function during post-training. Existing research attempts to mitigate overthinking from the perspective of prompt design or model training, but generally underestimates the importance of task difficulty awareness, which makes it difficult for LRMs to effectively allocate reasoning resources. In this paper, we propose Difficulty-aware Policy Optimization (DiPO), a reinforcement learning-based LRM training framework. DiPO encourages LRM to spontaneously model task complexity, and integrates them into reinforcement learning framework to adjust the generation preferences introduced by post-training. A difficulty modeling method based on model self-reasoning is proposed, which significantly reduces the dependence on manual annotation and formalize task complexity. We further develop a difficulty-signal-enhanced reward function that incorporates a penalty for lengthy reasoning while considering reasoning performance and output format. Experimental results indicate that DiPO enables the model to spontaneously adjust inference overhead, significantly reducing redundant tokens without losing performance due to thought compression.