Multimodal aspect-based sentiment analysis (MABSA), which includes aspect term extraction and aspect sentiment recognition based on available multimodal information, has recently attracted significant research attention. However, the research on MABSA thus far has typically failed to fully integrate different types of textual information inherent in an input sentence and overlooked the unalignment problem between different modalities. In addition, coarse-grained image filtering may inadvertently discard informative auxiliary visual cues, thereby limiting the performance of the model. To alleviate the aforementioned problems, we propose a multimodal method enhanced with Semantic-syntactic Graph Balancing and dynamic Image Selection (SGBIS) to improve performance in MABSA tasks. SGBIS consists mainly of three modules: (1) a semantic-syntactic graph balancing module, which constructs and adaptively fuses semantic and syntactic dependency graphs to enrich textual representations; (2) a dynamic image selection module, which calculates intra-image saliency scores and inter-modal correlation scores between image patches and textual tokens, and then employs the Gumbel-Softmax technique to select the most informative image patches while aggregating unselected patches as complementary visual cues; and (3) a cross-modal alignment and interaction module, which utilizes a learnable alignment matrix to facilitate fine-grained alignment and interactions between modalities. Comprehensive evaluations on two popular datasets reveal that SGBIS outperforms competitive baselines, validating its advantages in effectively integrating multimodal signals and enhancing performance in joint MABSA tasks.

Visual questions are an important means to evaluate students’ knowledge. Knowledge-based visual question classification can effectively excavate the knowledge intention of the question, and realize the effective organization and management of online question resources at the knowledge level. The existing methods simply regard it as a multimodal classification task, ignoring the capture of implicit knowledge information and the fine-grained interactions between multimodal and multi-granularity features. To mitigate this, we propose a quaternion hypergraph consistent network (QHCN). This approach can not only extract explicit semantic features and implicit knowledge features from text and images simultaneously, but also considers three key properties among explicit-implicit features: modality complementation, modality independence, and knowledge consistency. Specifically, a visual question is represented as a quaternion vector consisting of two modalities and four-dimensional features. To achieve multimodal complementation, the consistency of vision and language guides the construction of a quaternion hypergraph, and a quaternion convolution operator deeply fuses explicit-implicit features. To capture inter-dependencies between explicit-implicit features, the independence loss and knowledge consistency loss are designed to optimize hypergraph network parameters and enhance the hypergraph structure. Extensive experiments on visual question sets verify that our QHCN achieved an accuracy of 94.82% and an F1 score of 94.76%, outperforming the optimal baseline by +1.46% and +1.53%, respectively.

Collaborative problem-solving (CPS) activities encourage learners to share knowledge and complete complex tasks together. However, learners with insufficient prior knowledge may struggle to contribute effectively to these tasks. Appropriate support is therefore needed to help learners address challenges that arise during CPS. This study developed an AI personal assistant (AIPA) embedded with a ChatGPT-based large language model (LLM) to serve as a personal consultant in CPS activities by providing personalized knowledge, resources, and guidance. This design raises an important question about how AI assistance is associated with learners' interaction patterns and the quality of collaboration. By analyzing student dialogues and interactions with the AIPA, this study found no significant differences in CPS performance among high-, mixed-, and low-knowledge groups after using the AIPA. However, both mixed- and low-knowledge groups interacted with the AIPA more frequently than the high-knowledge groups. Analysis of question types further showed significant differences, with explanation-oriented questions being more common in low-knowledge groups. In addition, high-knowledge groups interacted minimally with the AIPA, low-knowledge groups engaged consistently, and mix- knowledge groups showed moderate interaction, with shifting question types across the CPS process. The results also showed that groups with lower prior knowledge relied more on information provided by the AIPA in their discussions. These findings suggest that the AIPA can serve as a supportive resource in CPS contexts and contribute to understanding students' interaction patterns with AI support in educational settings.

Background Recent research increasingly highlights the central role of interventions in enhancing shared monitoring during collaborative problem-solving. However, traditional intervention approaches suffer from limitations in timeliness and adaptability. Large language model (LLM), equipped with deep semantic parsing and contextual perception, can dynamically detect latent challenges and provide targeted, timely, context-sensitive feedback.Objectives This study examines how LLM-supported interventions affect group shared monitoring during dynamic CPS processes.Methods This study designed a collaborative problem-solving platform integrated with LLM, and 28 students from a university in China participated in CPS activities. Chi-square tests, conditional random fields, linear mixed models and correlation analyses were adopted to examine the changes in both monitoring behaviour and equality of monitoring participation in high-cohesion (HCGs) and low-cohesion groups (LCGs) after LLM-supported group metacognitive scaffolding (LLM-GMS) intervention, as well as their effects on collaborative performance.Results and Conclusions The results show that (1) LLM-GMS activated more socio-cognitive and behavioural monitoring in HCGs, whereas LCGs mainly exhibited heightened behavioural monitoring. (2) Descriptive analyses revealed divergent trends in monitoring participation equality across group types, with HCGs showing increased equality and LCGs exhibiting a decline. (3) In HCGs, socio-emotional monitoring was positively associated with collaborative performance, whereas participation equality and behavioural monitoring exhibited negative associations with collaborative performance. In contrast, among LCGs, behavioural monitoring was positively related to performance, whereas socio-cognitive monitoring was unexpectedly negatively associated with performance.Implications These findings highlight that LLM-GMS can be a valuable tool for supporting collaborative learning, but its effectiveness depends on group characteristics and its implementation approach.

The rise of micro-credentials and competency-based education demands a new credentialing paradigm that balances public trust with the pedagogical imperative of learner agency. Existing systems fail this balance, forcing a trade-off between transparency and privacy. This paper proposes a dual-chain architecture that resolves this conflict by merging a permissioned consortium blockchain for governed issuance with a public blockchain for open verification. Our framework integrates W3C Verifiable Credentials with the IEEE 3527.1texttrademark Digital Intelligence (DQ) standard for semantic interoperability and employs Zero-Knowledge Proofs (ZKPs) to empower learners with granular, privacy-preserving control over their data. Performance evaluation confirms the system’s robustness, with the consortium chain achieving 89.1 TPS and ZKPs verification completing in under 95 milliseconds. This work delivers a scalable technical solution and, in doing so, provides a practical foundation for a pedagogical model that enhances learner agency and supports self-regulated learning, paving the way for a more equitable and learner-centric educational ecosystem.

Multimodal emotion recognition in conversations (MERC) under uncertain missing modalities remains a critical challenge, especially in educational contexts. Significant research efforts have been dedicated to developing novel solutions to tackle this problem. Nevertheless, existing methods are difficult to fully utilize available information for high-quality data reconstruction, and fail to learn discriminative multimodal representations for handling uncertain missing modalities. Furthermore, the generalization of such methods in educational contexts remains largely overlooked. To address these challenges, this study proposes a novel graph fusion network approach for achieving MERC tasks in education with uncertain missing modalities. Our method first reconstructs missing modalities by using adaptive fusion weights to aggregate information from the most similar neighbors in the available modalities. The reconstructed data are processed via a mask-aware graph structure to explicitly exploit masked cues to enhance the robustness of intramodal representations. Initial multimodal fusion representations derived from single-modal features are employed to build a dual-graph structure for further refining the multimodal representation learning of each utterance. During the training process, our method also introduces the boundary-aware contrastive learning to enable the accurate identification of samples near decision boundaries. Experiments on three multimodal datasets, including one self-collected educational dataset, CMU-MOSI, and CMU-MOSEI, demonstrates its superiority, effectiveness, and generalizability.

With the accelerating growth of multimodal user-generated content, effectively aligning image and text representations to establish deep interactions has become a key challenge in Multimodal Aspect-Based Sentiment Analysis (MABSA). To address the limitations in existing methods, such as the lack of cross-modal structural alignment, the introduction of potential noise by simple semantic alignment, and shallow fusion strategies, we propose a novel Dual Perspective Hierarchical Alignment (DPHA) framework for joint MABSA. Specifically, DPHA leverages the denoised image descriptions and aligns them with the textual content at the structural level. To further enhance structural alignment, an unsupervised contrastive learning is proposed to minimize the distance between the dependency structures of image descriptions and the corresponding text. Meanwhile, DPHA performs a cascaded fusion of object regions and salient patches to make a balance between sufficient visual information and potential noise. The fused visual representations are then semantically aligned with textual information. Next, the alignment matrices computed from the two previous modules are used to guide attention flow across multiple Transformer layers, thereby achieving deep multimodal fusion. During training, a task-specific curriculum learning strategy is incorporated to enable dynamic sample scheduling to more effectively facilitate the acquisition of the model’s capability. Experiments on benchmark datasets have verified that DPHA significantly outperforms existing baselines. Ablation studies have also demonstrated the effectiveness of each component in the proposed DPHA.

The accurate recognition of students’ emotions and engagement levels in real classrooms remains a persistent challenge in Automated Emotion Recognition (AER), primarily due to limited accuracy, weak robustness under noise, and poor interpretability. In order to address the aforementioned issues, this paper proposes a dual-path multimodal framework that leverages Gemini 2.5 Pro through structured API interactions. The framework delineates two distinct strategies: The first pathway is referred to as the ‘End-to-End Multimodal Pathway’, and it integrates transcribed text and visual cues into Gemini for the purpose of achieving unified inference and reasoning. The second pathway is referred to as the ‘Decomposed Fusion Pathway’, and it processes text and vision inputs via separate prompt templates. It performs attention-based temporal fusion locally and then queries Gemini for pedagogical reasoning. Experiments on RAF-DB, FER-2013, and the real-world DIP+SEER dataset demonstrate that the End-to-End approach achieves higher accuracy and efficiency, while the Decomposed Fusion method offers improved robustness and interpretability. The present study demonstrates the potential of large multimodal models to enhance emotion and engagement recognition in authentic classroom environments and to support more adaptive, emotionaware educational interventions.

The rapid development of online education has contributed to an explosive growth of test question resources. To better organize test questions, test question annotation aims to assign relevant labels to test questions to support intelligent educational applications. However, the existing methods lack domain applicability and cannot accurately recognize labels of short test questions with sparse semantics and semantic confusion. To this end, we propose a semantic improved network model (QuesImNet) for test question representation, which enriches the semantics of questions by integrating the disciplinary context. Specifically, the discipline domain graph with knowledge attribute keywords and latent topics is constructed from the domain corpus and learns the task-oriented representation of disciplinary context via graph neural network. Moreover, a multi-view interaction with intra-interaction and inter-interaction attention mechanisms is designed to simulate the cross-semantics between the text words, topics and knowledge attribute keywords, thus effectively constraining their semantic representations. QuesImNet utilizes a series feature fusion strategy with dual switches, enabling flexible combinations of various modules to emphasize informative features and achieve optimal test question representation and annotation. To validate the effectiveness of QuesImNet, extensive experiments are conducted on a large-scale real-world math question dataset, the results show that QuesImNet has a high capability to enhance the question semantics and outperforms classical models in several question-based tasks.

Collaborative Problem Solving (CPS) is a multifaceted process where students with strong social and cognitive skills tend to excel. However, improving these skills requires an in-depth understanding of the relationship between skills and students' internal state of mind. It is difficult to assess a student's state of mind based on external cues like facial expressions, speech and behaviour, as these can be easily disguised. This study used a portable EEG device to measure brainwave signals during online CPS tasks, focusing on variations in brain activity using the Approximate Entropy (ApEn) metric to understand why students with high CPS skills perform better. Seventy-two undergraduates participated, with their brainwaves continuously recorded during CPS activities. The findings showed distinct brainwave patterns among students with different social and cognitive skills. Notably, brain activity was heightened during the problem-solving stage compared to the conceptualisation stage. Students with higher social skills showed significant activity in the alpha (inspiration), beta (decision-making) and gamma (creative thinking) bands, while those with stronger cognitive skills primarily activated the alpha band. These results highlight the need to tailor online CPS activities to effectively support students' development in both social and cognitive domains.

Knowledge graphs (KGs) are utilized to represent factual knowledge, constructing complete and accurate KGs is vital for precise knowledge representation. In mathematics education scenarios, well-defined KGs can significantly enhance both targeted instruction for teachers and comprehensive knowledge acquisition for learners. However, existing “text-to-graph” approaches face two main challenges: (1) generated KGs often suffer from either redundant or insufficient entities and relationships, resulting in structural imbalance and reduced pedagogical relevance, and (2) a lack of specialization for the unique requirements of mathematics education. To address these limitations, we propose a novel two-step semantic extraction framework implemented via an agent-based approach, specifically designed for mathematics education domain. Experimental results demonstrate that our method effectively enriches contextual information by aligning extractions with pedagogical dimensions, thereby facilitating more effective teaching and learning.

Emotion recognition in conversations has recently emerged as a hot research topic owing to its increasingly important role in developing intelligent empathy services. Thoroughly exploring the conversational context and accurately capturing emotion-shift information are highly crucial for accurate emotion recognition in conversations. However, existing studies generally failed to fully understand the complex conversational context due to their insufficient capabilities in extracting and integrating emotional cues. Moreover, they mainly focused on the speaker's emotion inertia while paying less attention to explore multi-perspective emotion-shift patterns. To address these limitations, this study proposes a novel multimodal approach, namely, GAT-CRESA (Graph ATtention based on Contextual Reasoning and Emotion-Shift Awareness). Specifically, the multi-turn global contextual reasoning module iteratively performs contextual perception and cognitive reasoning for efficiently understanding the global conversational context. Then, GAT-CRESA explores emotion-shift information among utterances from both the speaker-dependent and the global context-based perspectives. Next, the emotion-shift awareness graphs are constructed for extracting significant local-level conversational context, where edge relations are determined by the learnt emotion-shift labels. Finally, the outputs of graphs are concatenated for final emotion recognition. The loss of emotion prediction task is combined together with those of two perspective's emotion-shift learning for guiding the training process. Experimental results show that our GAT-CRESA achieves new state-of-art records with 72.77% ACC and 72.81% wa-F1 on IEMOCAP, and 65.44% ACC and 65.04% wa-F1 on MELD, respectively. The ablation results also indicate the effectiveness and rationality of each component in our approach.

Modern education aims at providing students with more personalized learning services and more engaging learning experiences. One promising approach is to develop educational agents to facilitate high-quality completion of various educational tasks. In recent years, the advent of large language models (LLMs) has breathed new life into educational agents and pushed them into a new stage of intelligence. This survey tries to conduct a comprehensive and thorough investigation of LLM-based agents in education. First, the developments of educational agents are presented as background information. Subsequently, we propose a unified architecture for LLM-based educational agents, including perception, profiling, memory, reasoning, and action modules, and summarize two primary methods (i.e., fine-tuning and prompt engineering) for equipping them with abilities. Next, we categorize the potential applications of LLM-based educational agents across the “teaching-learning-assessment-research” chain, and discover that LLM-based educational agent can play significant roles in various educational tasks. Furthermore, we reveal that when assessing the effectiveness of LLM-based educational agents, subjective evaluation remains dominant, supplemented by objective evaluation. Finally, the open issues and future research directions in this field are discussed from multiple perspectives. We hope that this survey can provide valuable insights and inspirations for researchers and practitioners to enhance the further development of educational agents in the future.

Socio-cognitive conflict (SCC) is an indispensable cognitive challenge and driving force in collaborative problem-solving, while socially shared regulation of learning (SSRL) is essential for addressing social challenges in collaboration. Although SCC has been classified in prior research, its dynamic process and impact on collaborative performance from an SSRL perspective remain underexplored. This study analyzes discourse data from 27 college students using regression, entropy, and association rule mining. Results show that recognition of anomaly conflict (RC) and anxiety conflict (AC) are positively associated with performance. Compared to low-performance groups (LPGs), high-performance groups (HPGs) in RC display more consistent SSRL behaviors and adopt a "planning - monitoring" optimization pattern. In interest conflict, HPGs follow a "reflection - monitoring" deepening pattern characterized by a dynamic cycle between reflective elaboration and behavioral monitoring. These findings underscore the importance of conflict management and offer practical implications for designing effective instructional interventions.

PurposeThis study aims to explore the potential of Multi-Agent Large Language Models (MALLM) to enhance Social Network Analysis (SNA) for online education. It compares MALLM with single-agent LLMs in conducting, interpreting and applying SNA, addressing barriers that limit adoption.Design/methodology/approachAn exploratory experiment using AutoGen compared MALLM and single-agent LLMs across multistep SNA workflows with a Coursera discussion data set. The process included data exploration, analysis and visualization. Specialized agent teams were assigned to analysis and interpretation. Performance was tested over 20 rounds, evaluated on comprehension, accuracy, execution and educational relevance.FindingsSingle agents were more efficient in simpler tasks (data exploration 85% vs 25%, visualization 50% vs 45%). MALLM outperformed in complex tasks, with higher SNA precision (30% vs 25%), stronger node-level analysis (95% vs 65%) and greater educational insights (55% vs 35%). However, MALLM faced coordination inefficiencies in linear tasks.Research limitations/implicationsLimitations include contextual forgetting, token-size constraints and coordination overhead. Results are specific to GPT-4/GPT-4o, with a 30% success rate in complex tasks, indicating LLMs are not yet sufficient for full automation.Practical implicationsMALLMs can advance online education by supporting personalized learning and engagement while democratizing access to advanced analytics and pedagogical feedback, thereby enhancing educational equity.Originality/valueTo the best of the authors' knowledge, this study is among the first to examine MALLMs' management of multimodal, domain-specific analytics tasks, moving beyond general text-based applications, highlighting their advantages in generating educational insights and informing agent design while providing benchmarks for advancing multi-agent LLM systems.

Sim-Stu is a multi-agent, retrieval-augmented learner simulation system that maps Self-Determination (SDS) and General Self-Efficacy (GSES) into computable decision parameters to generate knowledge-grounded learning processes. Standardized SDS/GSES scores drive a decision-execution agent to plan study steps, select resources, and produce auditable logs. A SDSxGSES design yields metrics across learning style, process, and path rationality. Within-profile variability is low to moderate, supporting inter-profile comparison. Main effects align with theory: higher GSES boosts persistence/affect and reduces help-seeking; higher SDS increases exploration/resource use. A nine-profile heatmap shows the highxhigh quadrant features strong retry behavior and higher completion. Sim-Stu provides a reproducible sandbox for pre-deployment instructional testing and learning analytics.

Blended learning has been widely integrated in college-level computer science education. Despite evidence about benefits of blended learning, students' in-class activities remain underexplored. To afford effective blended learning experience, supporting students in both modalities is essential. This study thus took an initial step to fill the gap by investigating college students' in-class activities in a blended course from the perspective of attention. Using non-intrusive electroencephalography (EEG) instruments to collect attentional data, this study found students' attention in in-class activities positively correlated with their learning gains. Students' attention also varied across in-class activities, reaching a higher level in group discussions than in pre-tests and lectures. Linear regression analysis indicated students' length of time spent viewing online resources and their pre-test scores significantly predicted their in-class attention. The findings of the study provide insight into course design and facilitation for effective blended computer science courses.

Purpose Although numerous signal modalities are available for emotion recognition, audio and visual modalities are the most common and predominant forms for human beings to express their emotional states in daily communication. Therefore, how to achieve automatic and accurate audiovisual emotion recognition is significantly important for developing engaging and empathetic human–computer interaction environment. However, two major challenges exist in the field of audiovisual emotion recognition: (1) how to effectively capture representations of each single modality and eliminate redundant features and (2) how to efficiently integrate information from these two modalities to generate discriminative representations. Design/methodology/approach A novel key-frame extraction-based attention fusion network (KE-AFN) is proposed for audiovisual emotion recognition. KE-AFN attempts to integrate key-frame extraction with multimodal interaction and fusion to enhance audiovisual representations and reduce redundant computation, filling the research gaps of existing approaches. Specifically, the local maximum–based content analysis is designed to extract key-frames from videos for the purpose of eliminating data redundancy. Two modules, including “Multi-head Attention-based Intra-modality Interaction Module” and “Multi-head Attention-based Cross-modality Interaction Module”, are proposed to mine and capture intra- and cross-modality interactions for further reducing data redundancy and producing more powerful multimodal representations. Findings Extensive experiments on two benchmark datasets (i.e. RAVDESS and CMU-MOSEI) demonstrate the effectiveness and rationality of KE-AFN. Specifically, (1) KE-AFN is superior to state-of-the-art baselines for audiovisual emotion recognition. (2) Exploring the supplementary and complementary information of different modalities can provide more emotional clues for better emotion recognition. (3) The proposed key-frame extraction strategy can enhance the performance by more than 2.79 per cent on accuracy. (4) Both exploring intra- and cross-modality interactions and employing attention-based audiovisual fusion can lead to better prediction performance. Originality/value The proposed KE-AFN can support the development of engaging and empathetic human–computer interaction environment.

Blended learning, integrating online and in-person components, has been increasingly adopted in higher education to enhance students’ learning experience and outcomes. While the advantages of blended learning are well-evidenced, research has primarily focused on the online pre-learning component, neglecting the significance of in-class activities. In-class activities play a crucial role in affording active learning opportunities (e.g., discussion, elaboration), necessitating a systemic understanding of their dynamics. The purpose of this study was thus to systemically investigate college students’ learning behaviors during in-class activities in a blended course. In-class activities were video-recorded and labelled manually following a coding scheme. By establishing a linear regression model, the study identified listening to the instructor’s lecture and taking notes as two predictors of students’ learning gains. Additionally, sequential patterns of learning behaviors during in-class activities were examined. The reciprocal interactions between students’ behavior of listening to the lecture and their note-taking actions were noted. The findings of this study contributed to a systemic view of blended learning by shedding light on students’ learning behaviors and their implications for instructional practice.

Collaborative problem-solving (CPS) involves the interaction and interdependence of students’ social and cognitive skills, making it a complex learning process. To delve into the complex dynamics of CPS, previous research has categorized socio-cognitive roles, providing insights into social-cognitive frameworks. However, despite the specific cognitive and social interaction structures employed by roles to engage in CPS interactions, most existing research primarily focuses on individual roles, neglecting inter-role interactions. To fill this gap, twelve triad groups were formed by engaging 36 undergraduate students in online CPS activities to examine differences in social and cognitive interaction structures across different roles and group compositions. Additionally, analyze the differences in CPS processes among various group compositions. The analyses identified five roles ( Lurkers, Followers, Drivers, Influential Actors , and Innovators ) and three group compositions ( Balanced groups, Decentralized groups , and Power Struggle groups ). The socio-cognitive structure of Balanced groups , along with other evidence, indicates effective information sharing and negotiation interactions. In contrast, Decentralized and Power Struggle groups exhibited various deficiencies in their socio-cognitive structures, negatively impacting group collaboration processes. These insights provide educators with a comprehensive guide to fostering effective group compositions and role dynamics in online CPS settings, thereby enhancing the overall success of CPS. Additionally, possible activity design considerations and scaffolding strategies are also discussed.