Understanding how humans process visual information during learning remains a central challenge in cognitive neuroscience. Traditional approaches typically rely on behavioral observations or indirect neural measures, which limits direct insight into learners' internal cognitive representations. Recent advances in brain--computer interfaces (BCIs) and deep learning have enabled image reconstruction from electroencephalography (EEG), providing a promising pathway to decode latent cognitive states. In this study, we investigate how EEG-based image reconstruction relates to students' learning. While students watched educational videos, we recorded EEG signals and used an image reconstruction model to generate images reflecting their perceived visual representations. We assessed reconstruction quality using semantic similarity between reconstructed and original images. First, we examine whether semantic similarity predicts overall academic performance and find a strong positive association. Second, we evaluate item-level effects and show that questions answered incorrectly correspond to significantly lower semantic similarity than correctly answered questions. Third, we analyze the temporal dynamics of reconstruction and demonstrate that temporal misalignment patterns---quantified through dispersion-related measures---are associated with learners' performance. Overall, these results suggest that EEG-based image reconstruction captures both the strength and the temporal evolution of learners' cognitive representations, offering a novel tool for studying learning mechanisms and supporting more personalized, data-driven educational interventions.

Emotion-cause pair extraction is a challenging task that focuses on identifying potential pairs of emotions and their corresponding causes within unannotated documents. Existing methods typically decompose this task into three sub-tasks, namely emotion clause extraction (EE), cause clause extraction (CE), and emotion-cause pair extraction (ECPE). However, these methods suffer from two limitations: (1) inadequacies in clause representations due to the constraints of anisotropic word embeddings; (2) a lack of comprehensive consideration of the document's structural information and the interactions between different sub-tasks. To address these issues, we propose a novel large language model augmented multi-task learning network (LLM-MTLN) with inter-clause modeling. Specifically, we fine-tune a large language model to realize the sufficient document encoding as well as to mitigate the problems associated with anisotropic word embeddings. In addition, the external sentiment knowledge is seamlessly integrated into the clauses for further enhancing the clause representations. Next, a directed graph is constructed to effectively model the inter-clause relationships in the document, so as to facilitate the transformation of causal information between clauses. On this basis, we tackle the EE, CE, and ECPE tasks jointly in a well-designed multi-task learning network, in which collaborative interactions and information sharing among the three sub-tasks are fully considered. The proposed approach is compared with 16 other baselines on two benchmark datasets in different languages. Extensive experimental results show that our proposed LLM-MTLN achieves new state-of-the-art performance in the ECPE task and exhibits strong adaptability to various linguistic environments. All code and data for this work is available at https://github.com/chaodongwen/LLM-MTLN.

Although large language models (LLMs) demonstrate significant potential for advancing personalized science education, they face challenges in generating science problem-solving processes adapted to students’ grade levels. In this paper, we developed a Chinese Science Question (CSQ) dataset, which comprises both a benchmark and a training set, aiming to evaluate and enhance the science problem-solving capabilities of LLMs. The CSQ consists of 12,000 high-quality samples featuring a variety of question types and diverse discipline properties, covering four subjects and multiple topics at the Chinese primary school. We further designed the language model to reflect these discipline properties in the generated responses, emulating the thought process of students when solving science questions. We demonstrated that CSQ and its extensive annotations can be employed for fine-tuning models. This was confirmed through both automatic and human evaluations, particularly in generating problem-solving processes that are aligned with students’ grade levels.

In the field of recommender systems, graph neural networks (GNNs) have been extensively applied to collaborative filtering to generate personalized recommendations for users. To solve the problem of lack of observed data and contrasting interactions during representation learning, graph contrastive learning as an effective self-supervised learning (SSL) technique is presented to obtain augmented user and item representations. Nevertheless, most self-supervised approaches to generate recommendation either disrupt the graph structure or node embeddings through random augmentations or introduce augmented SSL information from biased data through heuristic methods. To overcome these challenges, we propose a learnable view augmentation model for collaborative filtering (LACF). Specifically, our framework embeds parameterized learnable view generators layer by layer into the automatic augmentation strategy, thus dynamically optimizing the adaptive augmented views of users and items through the backpropagation of weight gradients. In addition, LACF introduces a multiscale learning strategy that guides the view generator with layer-wise aware optimization and graph-level adaptive augmentation, enabling joint learning of representations with topological heterogeneity and semantic similarity from integrated viewpoint, achieving superior view augmentation. Extensive experiments on realworld datasets demonstrate that our LACF outperforms state-of-the-art baselines. In-depth analysis confirms the advantages of LACF in resistance against noise disturbances, alleviating data sparsity, and improving training efficiency.

Junior high school English classroom questioning is currently plagued by overemphasis on low-order cognitive skills and lack of thinking gradients. Though Generative Artificial Intelligence (AIGC) offers a new solution, the instructional adaptability and effectiveness of LLM-generated questions need systematic verification. Through literature review, this study identifies three core indicators for evaluating question content effectiveness: cognitive level, content relevance and language appropriateness. Based on Bloom's Taxonomy of Educational Objectives and Krashen's Input Hypothesis, it constructs the Theory-to-Prompt Mapping (TPM) framework and develops the Doubao Classroom Questioning Agent accordingly. A comparative experiment is conducted to generate classroom questions, whose effectiveness is verified by lexical coverage analysis, FKGL measurement, LDA semantic analysis and expert evaluation. The results show the TPM-based agent can generate questions covering all six cognitive levels of Bloom's Taxonomy, with high alignment to textbook vocabulary and strong thematic relevance to textbook content. This study verifies the feasibility and effectiveness of the TPM framework-based agent in assisting junior high school English teachers to design effective classroom questions, and provides a replicable methodological paradigm for AI-enabled junior high school English classroom teaching.

As an innovative AI-driven chatbot, ChatGPT holds substantial potential to redefine humancomputer interaction and enhance teaching and learning. However, current applications often prioritize functional utility over active learning facilitation, leaving its role in promoting collaborative knowledge construction and metacognitive regulation empirically underexplored. To address this gap, this study integrated ChatGPT into pre-service teachers' collaborative learning experiences, leveraging the Six Thinking Hats thinking strategy to scaffold instructional design tasks. Using a mixed-methods approach combining co-regulation coding and epistemic network analysis (ENA), this study compared the co-regulation processes and focus patterns of groups using ChatGPT with the Six Thinking Hats strategy versus those using Six Thinking Hats alone. Results indicate that ChatGPT-assisted groups exhibited intensified co-regulatory focus on Evaluation, along with significantly heightened attention to Task Understanding and Content Monitoring. These findings not only illuminate ChatGPT's transformative role in teacher education but also provide actionable insights for educators designing AI-enhanced collaborative learning environments.

While Large Language Models (LLMs) are increasingly integrated into educational settings, the transition from passive chatbot to autonomous LLM-based agents—characterized by tool use, memory retention, and goal-directed reasoning—remains underexplored. To address this gap, this study designed a domain-specific educational agent (DBagent) for an undergraduate database course, and examined its impact on learning achievement, cognitive engagement patterns, and student perceptions. A large-scale quasi-experiment was conducted with 313 sophomore students across four authentic classes (three experimental and one control) over a four-week period. We employed statistical analyses to examine learning achievement and the distribution of cognitive engagement levels, lag sequential analysis (LSA) to analyze cognitive behavior sequences from interaction logs, and structural equation modeling (SEM) to interpret survey data. Findings indicate that (1) the agent-enriched environment significantly improved learning achievement compared with traditional instruction; (2) LSA revealed distinct cognitive trajectories: while student-instructor interactions fostered balanced cognitive levels, student-agent interactions were characterized by high-frequency engagement driven by a specific “Query-Evaluation-Query” verification loop; and (3) SEM analysis confirmed that learners’ positive perceptions of the educational agent promoted sustained engagement through the mediating role of satisfaction. Taken together, the findings indicate that educational agents can improve learning achievement, foster distinctive cognitive engagement patterns, and sustain learner participation through positive user perceptions. These findings provide empirical evidence for the pedagogical value of educational agents and offer practical implications for the design of GenAI-enriched learning environments.

This study investigated the factors shaping Chinese higher education students' continued intention and self-reported use of generative artificial intelligence (GenAI) tools in academic work. Drawing on an extended Unified Theory of Acceptance and Use of Technology (UTAUT) model, the study examined the roles of performance expectancy, effort expectancy, social influence, facilitating conditions, behavioural intention, use behaviour, and perceived knowledge. An explanatory sequential mixed-methods design was adopted. In the quantitative phase, survey data from 580 university students were analysed using structural equation modelling. In the qualitative phase, follow-up semi-structured interviews were conducted to explain and elaborate on the quantitative findings. The results showed that performance expectancy, social influence, and perceived knowledge significantly predicted students' behavioural intention to use GenAI, while facilitating conditions and behavioural intention were associated with self-reported use behaviour. Effort expectancy did not significantly predict behavioural intention, suggesting that ease of use alone may be insufficient to explain continued GenAI use among students who already perceive such tools as accessible. Interview findings further indicated that students used GenAI mainly for brainstorming, proofreading, explanation, and problem-solving, but they also expressed concerns about over-reliance, academic integrity, and the reliability of AI-generated content. The study contributes to GenAI continued-use research by showing that students' use of GenAI in academic contexts is shaped by expected usefulness, social influence, and their perceived knowledge of how to use and evaluate AI outputs. Practical implications are provided for designing AI literacy training, classroom guidance, and institutional policies for responsible GenAI use.

Cognitive engagement classification is an important task for analyzing learners' interactive discourse texts in online education. Highly accurate classification results can facilitate efficient instructional decisions and interventions in online course discussions. However, existing cognitive engagement classification models fail to achieve accurate and reliable results due to two primary challenges: (1) the inherent sparsity of discourse features and single-view representation limitations, and (2) insufficient performance in scenarios with limited labeled data. To address these challenges, this study proposes a Dual-view Cross Attention enhanced Semi-supervised Learning method (DCASL) for discourse cognitive engagement classification. A dual-view cross-attention mechanism is built to fuse cognitive psychological and generic semantic features to achieve complementary discourse representations to enhance the feature fusion depth and receptive field of the classification model. By combining the dual-view cross-attention mechanism with supervised training and consistency training, we develop a semi-supervised cognitive engagement classification model. The dual-view feature enhancement significantly improves classification accuracy and model performance, particularly in scenarios with scarce labeled data. The results of experiments on three datasets from real-world online course discussions with eleven benchmark models show that DCASL effectively enhances model representation, improves cognitive engagement classification accuracy and is interpretable in the case of a small annotated corpus.

With the aid of artificial intelligence (AI), it is more feasible to leverage discussion data to understand the online collaborative learning process. This paper presented an AI-supported discussion representational tool (integrating behavioral and cognitive representations) aimed at enhancing online collaborative learning from three aspects: motivation, cognitive presence, and learning performance. A randomized controlled trial (RCT) was conducted to examine the tool's effectiveness with 122 students in four groups: (1) behavioral representation (n = 31), (2) cognitive representation (n = 31), (3) mixed mode (combining behavioral and cognitive representations, n = 30), and (4) a control group (n = 30). Results indicated that: (1) the discussion representational tool did not significantly enhance students' motivation but led to significant gains in their learning performance compared to the control group; (2) students who learned with the discussion representational tool showed significant improvements in higher-order cognitive presence, ordered network analysis revealed that they generated more higher-level cognitive connections; (3) the motivation is an effective predictor of cognitive presence and learning performance, and discussion representational tool positively moderated the relationship between motivation, cognitive presence, and learning performance. These findings represent a new contribution of the AI-supported discussion representational tool to facilitate online collaborative learning.

Forum discussions in Massive Open Online Courses (MOOCs) play a crucial role in promoting learning engagement and academic achievement. In particular, discussion topics significantly influence learners’ emotional and cognitive engagement. However, the complex interrelationships among these factors remain underexplored. This study introduces an innovative two-step methodological approach to investigate the relationships between topic complexity, emotional engagement, cognitive engagement, and academic achievement in MOOC discussions. Using BERT for engagement detection and developing a Joint Emotion and Cognition Topic Model (JECTM) based on Bayesian networks, we analyzed 27,428 discussion posts from 2857 learners in a psychology MOOC. Our findings reveal three key insights: (1) The proposed two-step approach efficiently detects topics and analyzes their patterns of emotional and cognitive engagement. (2) As topic complexity increases, learners demonstrate higher-order cognitive engagement while experiencing reduced positive emotions along with increased confused and negative emotions. (3) In high-complexity topics, learners who maintain both positive emotions and higher-order cognitive engagement are more likely to achieve academic success than those who have negative emotions or lower-order cognitive engagement. These fine-grained analyses provide valuable insights for optimizing discussion design and interventions. This study also provides implications for the analysis of classroom dialogues and AI tutor-based conversations.

Socratic teaching, known for its emphasis on heuristic questioning and deep thinking, has demonstrated significant advantages in promoting students’ cognitive development. However, traditional Socratic teaching places high demands on teachers’ expertise and real-time feedback capabilities, making it difficult to scale in large educational settings. Recent breakthroughs in large language models (LLMs) in natural language generation and dialogue comprehension offer the potential for automated Socratic teaching. In this paper, we propose Knowledge-Enlightened Learning Enhanced by LLMs (KELE), a novel multi-agent framework for structured Socratic teaching with LLMs. KELE constructs a structured Socratic teaching rule system (SocRule) and a “consultant–teacher” multi-agent collaborative teaching mechanism, in which two LLMs respectively take charge of teaching planning and execution, ensuring a logically coherent and hierarchically structured Socratic teaching process. We also construct SocratDataset, a structured Socratic teaching dataset covering 34 teaching strategies and over 42,000 dialogue turns, and train SocratTeachLLM, a specialized LLM for Socratic teaching tasks. Additionally, we build a comprehensive Socratic teaching quality evaluation system for LLMs, covering 9 dimensions from single-turn dialogue to multi-turn teaching processes. Experimental results show that SocratTeachLLM significantly outperforms GPT-4o, which has a much larger parameter size, across all Socratic teaching capabilities.

In the field of recommender systems, self-supervised learning has become an effective framework. In response to the noisy interaction behaviors in realworld scenarios, as well as the skewed distribution influenced by data sparsity and popularity bias, graph contrastive learning has been introduced as a powerful self-supervised method in collaborative filtering (CF) to learn enhanced user and item representations. Despite their success, neither heuristic manual enhancement methods nor the use of final node representations to construct contrastive pairs are sufficient to provide effective and rich self-supervised signals to regulate the training process. Therefore, the learned representations of users and items are either fragile or lack heuristic guidance. In light of this, we propose the Hierarchical multiview graph contrastive learning framework HMCF, which leverages the message passing mechanism at the layer level to introduce different granularity levels of view augmentation using supervised signals, thus better enhancing the CF paradigm. HMCF leverages rich, high-quality self-supervised signals from different granularity views for accurate contrastive optimization, helping to alleviate data sparsity and noise issues. It also explains the hierarchical topology and relative distances between nodes in the original graph. Comprehensive experiments on three public datasets shows that our model significantly outperforms the state-of-the-art baselines.

Various machine learning approaches have gained significant popularity for the automated classification of educational text to identify indicators of learning engagement -- i.e. learning engagement classification (LEC). LEC can offer comprehensive insights into human learning processes, attracting significant interest from diverse research communities, including Natural Language Processing (NLP), Learning Analytics, and Educational Data Mining. Recently, Large Language Models (LLMs), such as ChatGPT, have demonstrated remarkable performance in various NLP tasks. However, their comprehensive evaluation and improvement approaches in LEC tasks have not been thoroughly investigated. In this study, we propose the Annotation Guidelines-based Knowledge Augmentation (AGKA) approach to improve LLMs. AGKA employs GPT 4.0 to retrieve label definition knowledge from annotation guidelines, and then applies the random under-sampler to select a few typical examples. Subsequently, we conduct a systematic evaluation benchmark of LEC, which includes six LEC datasets covering behavior classification (question and urgency level), emotion classification (binary and epistemic emotion), and cognition classification (opinion and cognitive presence). The study results demonstrate that AGKA can enhance non-fine-tuned LLMs, particularly GPT 4.0 and Llama 3 70B. GPT 4.0 with AGKA few-shot outperforms full-shot fine-tuned models such as BERT and RoBERTa on simple binary classification datasets. However, GPT 4.0 lags in multi-class tasks that require a deep understanding of complex semantic information. Notably, Llama 3 70B with AGKA is a promising combination based on open-source LLM, because its performance is on par with closed-source GPT 4.0 with AGKA. In addition, LLMs struggle to distinguish between labels with similar names in multi-class classification.

Academic stress is a common psychological issue among college students, but the mechanism by which it affects academic performance is not well understood. This study extends existing literature by exploring a chain-mediated model composed of academic stress, learning motivation, depression, and academic performance. The study involved 591 students from a university in Hubei Province, China, and employed structural equation modelling to analyse the data. The results showed that academic stress did not have a significant direct effect on academic performance. Instead, depression played a significant mediating role between academic stress and academic performance. Further, learning motivation also acted as a mediator in the cascade between academic stress, depression, and academic performance. Subsequently, multi-group structural equation modelling was applied to analyse the data, revealing similarities and differences in the associations between variables based on sex. The results indicated that sex played a moderating role in the relationships among academic stress, depression, learning motivation, and academic performance. The implications of these findings were discussed.

Dialogue datasets are essential for advancing natural language processing (NLP) tasks. However, many existing datasets lack integrated annotations for personality and emotion, limiting models’ ability to effectively capture these aspects and generate personalized, human-like dialogues, which ultimately impact user experience. To address this challenge, we construct bilingual dialogue datasets in Chinese and English, incorporating Big Five personality traits and emotion annotations. We utilize the AutoGen tool within a multi-agent framework to generate multi-turn question-answering dialogue datasets based on fables. By creating persona agents with diverse personalities, we effectively enhance the heterogeneity of personalities, overcoming previous limitations in personality diversity. Finally, we validate the utterance quality in the dataset and investigate the alignment between conversational utterances and speakers’ personality traits. Moreover, by integrating emotional annotations for each utterance, This dataset offers significant potential for developing emotion-aware systems that automatically detect personality traits. It serves as a valuable resource for advancing emotionally intelligent dialogue systems and research in personality and affective computing.