Exploring the Interplay of Topic Complexity,Emotional Engagement,and Cognitive Engagement in MOOC Discussions: Using Deep Learning and Topic Modeling

Abstract

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.

Keywords

MOOCs topic complexity emotional engagement cognitive engagement deep learning topic modeling forum discussion

Introduction

Massive Open Online Courses (MOOCs) have revolutionized education by offering unprecedented global learning opportunities (Xing & Du, 2023). The scale of this transformation is remarkable - as of July 2023, Udemy alone serves 662 million learners across 202,000 courses (Dhawal & Archisha, 2023). However, MOOCs continue to face significant challenges in sustaining learner engagement, primarily due to limited face-to-face interaction and potential feelings of isolation in online learning environments (Orji & Vassileva, 2023; Tang et al., 2019). To address these challenges, online discussion forums have emerged as a vital solution, providing learners with opportunities to connect, share ideas, and support each other in ways that effectively mirror traditional classroom interactions (Gašević et al., 2018).

Research has consistently demonstrated that well-designed forum discussions play a crucial role in enhancing both learner engagement and academic achievement (Liu et al., 2024; Ouyang, 2021; Ramirez-Arellano et al., 2019). The effectiveness of these discussions is largely determined by topic design (Chen et al., 2021), particularly topic complexity - defined as the level of cognitive effort required for meaningful participation (Bloom, 1956; Nistor et al., 2020). This concept is firmly grounded in established frameworks such as Bloom’s taxonomy, which posits that analysis-level topics require more higher-order knowledge transfer compared to recall-level topics (Krathwohl, 2002). We randomly selected two discussion threads from our dataset to illustrate possible emotional and cognitive responses under different topic complexity levels. As illustrated in Figure 1, the discussion threads show distinct patterns of emotional and cognitive engagement. Thread (a) represents a low-complexity topic focused on understanding, which tends to elicit positive emotion and lower-order cognitive engagement. In contrast, thread (b) represents a high-complexity topic focused on creation, which tends to trigger confusion and higher-order cognitive engagement. These examples demonstrate hypothetical relationships between topic complexity and engagement patterns that require further empirical investigation to validate.

Learning engagement in online education encompasses behavioral, emotional, and cognitive dimensions (D’Mello, Dieterle, & Duckworth, 2017; Sun & Rueda, 2012). While researchers have extensively examined how topic complexity affects behavioral engagement (Chen et al., 2020; Lee & Recker, 2021; Peng & Xu, 2020), we know less about its simultaneous impact on emotional and cognitive engagement. This knowledge gap warrants attention for several key reasons.

First, emotional and cognitive engagement provide deeper insights into learning processes compared to surface-level behavioral engagement like participation rates and time spent on tasks (Liu et al., 2024). Specifically, cognitive engagement directly reflects learners’ mental investment by revealing how they analyze, synthesize, and construct knowledge (Chi & Wylie, 2014). Emotional engagement reflects learners’ motivation, persistence, and willingness to engage with course content (Dubovi, 2022). By understanding how topic complexity influences these engagement patterns, we can uncover crucial mechanisms for optimizing learning outcomes in MOOCs.

Second, topic complexity influences emotional and cognitive engagement in mixed ways. Complex topics may promote higher-order thinking while inducing negative emotions such as confusion or frustration (Wei et al., 2021). In contrast, simpler topics may elicit positive emotional responses but provide insufficient cognitive challenge. Understanding this delicate balance is essential for maintaining both positive emotional states and deep cognitive processing.

Third, we lack a comprehensive understanding of how topic complexity, emotional engagement, and cognitive engagement collectively impact academic achievement (Dillon et al., 2016; Wei et al., 2021; Yang et al., 2022). This gap offers an opportunity to develop an integrated model explaining how these factors work together to promote learning success in MOOCs.

In addition, analyzing the discussion topics and their influence on emotional and cognitive engagement in large-scale discussions poses substantial methodological challenges. For example, current methods for analyzing discussions typically rely on isolated engagement classification or topic detection. These isolated approaches have proven insufficient to accurately capture patterns of engagement and to model the complex interactions between these multifaceted factors (Dalipi et al., 2021). This critical gap requires innovative methodological solutions capable of conducting analysis of MOOC forum discussions at scale. To address these theoretical and methodological challenges, this study aims to:

1) Develop a novel two-step approach involving a deep neural network and a Bayesian probabilistic topic model to identify latent triggering relationships between topic complexity, emotional engagement, and cognitive engagement.

2) Investigate the interplay relationships between topic complexity, emotional engagement, and cognitive engagement and how they jointly predict academic achievement. These findings will provide valuable insights for optimizing the design and intervention strategies in MOOC forum discussions.

Literature Review

Topic Complexity in Discussions

Vygotsky’s social constructivist theory posits that knowledge is shaped, co-constructed, and learned within a knowledge community through language (Vygotsky, 1978). Lave & Wenger (1991) define learning communities as “the development of a shared identity around a set of challenging topics,” emphasizing the importance of discussion topics in fostering meaningful interactions. The complexity level of a discussion topic refers to the degree of effort and cognitive processing required in a collaborative discussion activity. High-complexity topics demand more cooperation, information extraction, and information synthesis from learners (Malinen, 2015).

Bloom’s taxonomy of educational objectives provides a widely used framework for analyzing topic complexity (Bloom, 1956; Krathwohl, 2002; O’Riordan et al., 2021). Building on this foundation, Ertmer et al. (2011) developed cognitive key verbs aligned with specific cognitive objectives to classify discussion topics into low-complexity (knowledge, comprehension, and application) and high-complexity topics (analysis, synthesis, and evaluation). Recent studies by Mohammed and Omar (2020) and Shaikh et al. (2021) have combined Bloom’s taxonomy with machine learning techniques to further refine topic complexity detection. They found that assessing topic complexity requires considering both cognitive key verbs and the difficulty of knowledge terms within the topic. These theoretical frameworks have been increasingly applied in computational approaches to analyze the topic complexity of discussions. For example, Iqbal et al. (2023) used Bloom’s taxonomy to categorize topics mined by Latent Dirichlet Allocation (LDA). Rivera et al. (2024) proposed a four-level classification scheme of discussion topics: asking questions, expressing opinions, analyzing content, and creating materials. This classification scheme offers a more nuanced approach to understanding the cognitive demands of different discussion topics. Chen et al. (2020) focused on evaluation and knowledge content in their topic complexity classification, highlighting the topics of distinguishing between factual recall and critical analysis in online discussions.

While multiple factors, including instructor presence, platform design, and course structure, contribute to learning engagement in MOOC discussions (Liu et al., 2024; Rivera et al., 2024), topic complexity emerges as a key focus for both practical and theoretical reasons. From a practical perspective, topic complexity serves three essential functions: first, it acts as a fundamental scaffold that structures and guides meaningful discussion (Chen et al., 2020); second, it represents a uniquely controllable and measurable variable that instructors can manipulate; third, it exhibits remarkable consistency across diverse discussion environments in online learning platforms. Furthermore, topic complexity has strong theoretical foundations and provides robust evidence of its pivotal role in shaping both cognitive processing and emotional states (Wei et al., 2021). These reasons make topic complexity an ideal factor for investigating the research objectives of this study.

Emotional Engagement and Cognitive Engagement in Discussions

Learning engagement in forum discussions encompasses behavioral, emotional, and cognitive dimensions that predict academic achievement (Sun & Rueda, 2012). Behavioral engagement refers to the duration and frequency of learning behaviors, and the role of topic complexity on behavioral engagement has been widely studied (Chen et al., 2020; Lee & Recker, 2021; Peng & Xu, 2020). Cognitive engagement reflects deeper levels of mental processing such as analyzing, synthesizing, and constructing knowledge (Chi & Wylie, 2014). Emotional engagement refers to learners’ affective responses during learning (Demir et al., 2023; Sun & Rueda, 2012). Topic complexity can lead to cognitive imbalance and affect learners’ sense of control over the learning process, resulting in a wide range of emotional responses (Buhr et al., 2019).

This study examines emotional engagement through the framework of epistemic emotions, which are intrinsically connected to cognitive processes and extensively studied in educational research (Han et al., 2021; Liu et al., 2022). Epistemic emotions function as key facilitators of cognitive processes, encompassing positive emotions (e.g., enjoyment), confusion (e.g., doubt), and negative emotions (e.g., frustration) (Chevrier et al., 2019). The phenomenon of confusion in learning represents a particularly nuanced case, as it occupies an intermediate position between positive and negative emotional states (D’Mello et al., 2014). When learners successfully resolve their confusion, it often leads to positive emotions, whereas unresolved confusion tends to generate negative emotions. Research indicates that confusion is a common experience during learning, with studies documenting its prevalence in 15%–50% of learning processes (Dillon et al., 2016). Based on these theoretical and empirical foundations, and following recent MOOC forum studies (Liu et al., 2022; Liu et al., 2024), the present study analyzes emotional engagement in discussions across three primary dimensions: positive, confused, and negative.

Cognitive engagement refers to the depth of mental processing that occurs when learners interact with learning materials. This study employs the Interactive, Constructive, Active, and Passive (ICAP) framework to measure cognitive engagement (Chi & Wylie, 2014). Numerous studies have substantiated the effectiveness of this framework (Atapattu et al., 2019; Liu et al., 2022; Wang et al., 2016). The ICAP framework posits that cognitive engagement and its contribution to academic achievement progressively increase from passive to active, constructive, and interactive cognition. Passive engagement involves merely acquiring information, which is not typically observable in discussion processes. Active engagement entails extracting information from existing knowledge. Constructive engagement indicates the integration of new knowledge with prior understanding. Interactive engagement represents a co-inferring process where learners collaboratively generate knowledge. Liu et al. (2022) further refined the coding scheme developed by Wang et al. (2015) based on the ICAP framework and applied it to analyze MOOC discussions, providing a robust methodology for assessing cognitive engagement in online learning environments.

Recent research suggests that topic complexity simultaneously influences learners’ cognitive and emotional engagement (Geng et al., 2020). According to cognitive load theory (Merriënboer & Sweller, 2005), topic complexity directly affects intrinsic cognitive load - low-complexity topics may induce boredom and lower-order cognitive engagement due to insufficient cognitive challenge, while highly complex topics may lead to frustration and excessive cognitive load that overwhelms working memory capacity, potentially limiting higher-order cognitive engagement (Ding, 2019; O’Riordan et al., 2021). However, the interplay between topic complexity, emotional and cognitive engagement, and their subsequent impact on academic achievement in online discussion remains to be fully explored.

Automated Analytics of Topic, Emotion, and Cognition in Discussions

The manual annotation of massive discussion data is both costly and impractical. The development of natural language processing (NLP) provides promising solutions for automated analysis of forum discussions. Recently, Bidirectional Encoder Representations from Transformers (BERT) have dramatically improved the performance of automated detection of emotion and cognition (Han et al., 2021; Liu et al., 2022; Liu et al., 2024). These findings highlight the robustness and efficiency of BERT in analyzing complex linguistic patterns in educational contexts. For example, Liu et al. (2022) employed BERT to detect emotional and cognitive engagement, observing enhancements in F1 scores by 18% and 8%, respectively, in comparison to text classification models based on machine learning, such as random forest (RF) and support vector machine (SVM). Although large language models (LLMs) such as GPT-4 and LLaMA-3 perform well with zero or few data samples, several evaluation studies have shown that BERT with sufficient data samples (>1000 samples) significantly outperforms LLMs in multi-class text classification tasks of emotion and cognition (Hou et al., 2024; Liu et al., 2024; Wang et al., 2023).

In addition, analyzing discussion using only emotional or cognitive indicators without detecting their co-occurring topics may hinder a refined understanding of the discussion process (Pan et al., 2022). LDA is an unsupervised machine learning algorithm based on the Bayesian network model for topic modeling (Blei et al., 2003), allowing automatic extraction of latent topics from discussion texts and coding their complexity. For example, Gašević et al. (2018) used LDA to identify topics and code them into critical thinking, social cohesion, and course logistics. Onan and Toçoğlu (2020) employed LDA to cluster discussion questions into three topics: technique, course content, and course logistics. These applications demonstrate the versatility of LDA in uncovering meaningful patterns in educational discourse.

Moreover, as a Bayesian network, LDA has good interpretability and a complete mathematical foundation. Thus, fusing LDA with feature variables (e.g., emotion or cognition) can effectively model complex discussion structures, providing a comprehensive framework for analyzing the multifaceted nature of online learning interactions. For example, Qiu et al. (2013) introduced a behavioral topic model (B-LDA) that jointly models discussion topic content and behavior patterns. Peng et al. (2020) developed an emotion and behavior topic model with an emotion lexicon to calculate emotional and behavioral patterns of topics, finding that course content topics elicited positive emotions, whereas course logistics topics triggered negative emotions. These studies highlight the potential of integrating emotional and behavioral factors into topic modeling for a more holistic understanding of online discussions.

To achieve our research objectives, after careful consideration of alternative methodologies, we determined that a single-step approach would be insufficient for two main reasons. First, while deep learning-based text classification algorithms excel at detecting emotions and cognitions, they lack the ability to effectively uncover latent topic structures within the text. Second, traditional topic modeling approaches are limited in modeling the intricate relationships between topics, emotions, and cognitions, resulting in reduced accuracy for both engagement detection and topic discovery (Birjali et al., 2021).

To address these limitations, we developed an integrated two-step approach that synergistically combines deep learning text classification with advanced topic modeling. This hybrid methodology effectively fulfills our research requirements through two steps: Step 1 leverages BERT’s powerful capabilities to achieve high-precision classification of emotional and cognitive engagement, while Step 2 uses a model to explore the complex relationships between topics, emotions, and cognition. Together, these components enable the discovery of meaningful and interpretable patterns in how engagement varies across topics.

Research Questions

To achieve the research objectives, we propose a two-step approach to investigate the following research questions (RQs):

RQ1: What are the discussion topics and their levels of complexity in the MOOC forum?

RQ2: How does topic complexity influence emotional and cognitive engagement in MOOC discussions?

RQ2a: How does the topic complexity influence emotional engagement?

RQ2b: How does the topic complexity influence cognitive engagement?

RQ2c: How does the topic complexity jointly influence emotional and cognitive engagement?

RQ3: How do topic complexity, emotional engagement, and cognitive engagement predict academic achievement?

Methods

Research Design

This study employs an evidence-centered learning analytic framework (see Figure 2) that systematically maps observable textual data to meaningful educational psychology indicators (Gašević et al., 2017; Wise et al., 2021). The research design comprises three distinct phases with six main steps: Phase A focuses on data collection, Phase B involves detecting meaningful indicators from discussion texts, and Phase C analyzes relationships between these indicators.

Phase A implements two key steps: (1) instructors design learning activities to facilitate MOOC forum discussions; (2) a comprehensive assessment program measures learners’ academic achievement. Phase B consists of three steps: (3) trained annotators label emotional and cognitive engagement in discussion text subsets to create training datasets for two BERT-based classification models that analyze unlabeled discussions; (4) the Joint Emotion and Cognition Topic Model (JECTM) detects implicit topics and their relationships with emotional and cognitive engagement using discussion posts’ engagement labels and word tokens; (5) two annotators code topic complexity levels. Phase C involves (6) applying statistical analyses to examine relationships among the research variables. Figure 2 shows a processed data sample for analysis at the bottom.

Instructional Design and Achievement Measurement

In steps 1 and 2 of Figure 2, experimental data was collected from a national-level quality course titled Modern Psychology. The study included 2857 Chinese undergraduate learners majoring in non-psychology fields (e.g., management and computer science), all with similar levels of psychological knowledge and undisclosed gender information. A team of psychological instructors from a university designed and conducted the course from May to September 2019, hosting it on China’s largest MOOC platform (https://www.icourse163.org). This nine-week course comprised nine units featuring lecture videos, course materials, and forum discussions. Instructors and learners contributed 523 root posts with varying topic complexity, with learners encouraged to express their viewpoints on each post.

The selection of Modern Psychology, a general education introductory course, was motivated by two considerations. First, the course encompasses diverse and balanced discussion topics of varying complexity levels, integrating conceptual knowledge comprehension, real-world case analysis, and innovative solution development. This range of topics provides an optimal environment for comprehensively examining the research questions. Second, psychology’s inherently interdisciplinary nature encompasses multifaceted content, which substantially enhances the generalizability of our research findings across multiple academic domains, such as statistics, neuroscience, and social psychology.

The achievement assessment, totaling 100 points, consisted of two components: lecture video viewing duration and quiz performance (50 points) and a final exam (50 points). Passing certificates were awarded to 1153 learners (40.36%) who scored 60 or above. Conversely, 1704 learners (59.64%) did not receive a certificate, as their scores were below 60. We consider the learners who received a passing certificate as the high-achievement group and those who did not receive a certificate as the low-achievement group. Many studies have adopted this approach to measuring academic achievement and dividing learners into high and low achievement groups (Doo & Kim, 2024; Liu et al., 2022; Liu et al., 2024). The certification process for this course is free, resulting in a higher certification pass rate than courses offered on paid platforms such as Coursera and EdX, effectively reducing the impact of financial factors on measures of academic achievement. In addition, forum participation was entirely voluntary with no mandatory requirements or course credits, ensuring that participants’ contributions stemmed from intrinsic interest rather than external motivations.

To optimize our topic modeling quality, we implemented a content word retention strategy that effectively reduced textual noise. This strategy has been extensively validated through multiple studies (Peng et al., 2020; Yao et al., 2017). We utilized HanLP’s part-of-speech tagging functionality to identify essential content words (including nouns, verbs, and named entities) while filtering out low-frequency words (occurring fewer than 5 times) (He, 2021). Subsequently, posts containing less than 3 words were flagged as spam because they could not be used for topic modeling. Our analysis encompassed 27,502 learner discussion posts, of which 27,428 were determined to be valid contributions, with only 74 posts classified as invalid. Notably, the average number of posts by the high-achievement group (mean = 11.15, SD = 8.24) significantly surpassed that of the low-achievement group (mean = 3.30, SD = 3.87).

Detecting Emotional and Cognitive Engagement Using BERT

As illustrated in step 3 of Figure 2, we employed a validated coding scheme to code emotional and cognitive engagement in the training and testing dataset (see Table 1) (Liu et al., 2022). Subsequently, this annotated dataset served as the foundation for training two BERT-based text classification models to detect emotional and cognitive patterns in unlabeled discussion posts.

Table 1.

The Coding Scheme of Emotional Engagement and Cognitive Engagement (S. Liu et al., 2022).

Engagement	Indicator	Description	Example
Cognitive engagement	Active	A learner refers to or explains concepts and ideas from the course materials. It is considered a lower-order cognitive engagement.	心理学是一门研究人类心理和行为的学科。/ Psychology is a subject for researching the mentality and behavior of human beings.
	Constructive	A learner constructs new ideas based on concepts present in the course material. It is considered a middle-order cognitive engagement.	我认为人类有追求幸福和快乐的本能,这可能是选择性失忆的一个原因。/ I think humans have an instinct to pursue happiness and pleasure, which may be a reason for selective amnesia
	Interactive	A learner constructs new ideas or poses new questions based on their peers’ responses. It is considered a higher-order cognitive engagement.	津巴多将学习定义为一种心理过程,而不是一种行为。你认为这两个词之间有区别吗?/ Zimbardo defined learning as a mental process rather than a behavior. Do you think there is a difference between these two terms?
Emotional engagement	Positive	A learner expresses positive emotions, such as curiosity, pride, enjoyment, surprise, calm, and hope.	感谢您基于积极心理学提出的建议,这对我的生活帮助很大。/ Thank you for your advice based on positive psychology. It helps me a lot in my life.
	Confusion	A learner expresses confused emotions, for example, doubt, conjecture, suspicion, and puzzlement.	您如何进行自我调节以完成困难但重要的任务?/ How do you self-regulate to accomplish difficult but important tasks?
	Negative	A learner expresses negative emotions such as boredom, frustration, anxiety, hopelessness, anger, and shame.	这门心理学课程相当难,因为第一周的课程内容让我感到压力很大。/ This psychology course was quite difficult because the course content in the first week made me feel pressured.

To establish the reliability of the coding scheme, two trained annotators independently annotated a set of 300 posts. The annotations of the two annotators are in favorable coherence. In particular, cognitive engagement achieved a Cohen’s kappa of 0.76, while emotional engagement reached a Cohen’s kappa of 0.82, exceeding the threshold of 0.70 (Cohen, 1960). Subsequently, the two annotators annotated 6000 posts as training datasets, as summarized in Table 2.

Table 2.

Sample Distribution of Labeled Datasets of Emotional and Cognitive Engagement.

Engagement	Indicator	N	%
Emotional engagement	Positive	2864	47.70
	Confusion	2095	34.90
	Negative	1041	17.30
Cognitive engagement	Interactive	1578	26.30
	Constructive	3162	52.70
	Active	1260	21.00

For automatic engagement detection, we trained and tested the BERT-based classification models using 10-fold cross-validation. BERT has demonstrated superior performance over traditional approaches like random forests, CNNs, and LSTMs in analyzing educational forum discussions (Devlin et al., 2019; Kong et al., 2025; Liu et al., 2022). We utilized the bert-base-Chinese model architecture (12 layers, 768 hidden dimensions, 12 attention heads) and fine-tuned it with optimized hyperparameters: 2 × 10⁻⁵ learning rate with linear decay, 32 batch size, 512 maximum sequence length, and 50 training epochs. The model training was conducted on a single NVIDIA V100 GPU utilizing mixed precision computation.

The classification performance of BERT demonstrated exceptional performance across both engagement dimensions (macro-F1 > 0.70) (Kong et al., 2025; Zou et al., 2021). For emotional engagement detection, the model achieved strong metrics with 0.837 accuracy, 0.793 macro-precision, 0.776 macro-recall, and 0.778 macro-F1 score. Similarly, robust performance was observed for cognitive engagement classification, yielding 0.754 accuracy, 0.758 macro-precision, 0.754 macro-recall, and 0.748 macro-F1 score.

Modeling Latent Relationships Between Topic, Emotion, and Cognition by JECTM

In step 4 of Figure 2, a Bayesian network model called JECTM was developed to extract latent topics that trigger emotional and cognitive engagement from large-scale discussion texts. As shown in Figure 3, the model incorporated emotional and cognitive features detected from BERT as heuristic information to guide the topic-generation process.

Figure 3 illustrates JECTM’s three-layer structure: learners (L), documents/posts (D), and words (W), with arrows indicating triggering relationships between variables. The model’s generation process follows these assumptions: (1) When a learner writes a post, a topic $z_{t}$ is drawn from the learner’s topic multi-distribution $θ_{l} = {z_{1}, z_{2}, \dots, z_{t}} (1 \leq t \leq T)$ . (2) Each topic $z_{t}$ has three multi-distributions of words $ϕ_{t}$ , emotions $π_{t e}$ , and cognitions $ψ_{t c}$ . (3) A post contains W words drawn from $ϕ_{t}$ , an emotion e drawn from $π_{t e}$ , and a cognition c drawn from $ψ_{t c}$ . The emotion e and cognition c were recognized using the BERT models. Detailed Equations (A1–A6) and Algorithm A1 are provided in the Appendix. The code is available at: https://github.com/hellodannyliu/JECTM.

JECTM’s design was guided by three key principles. First, we aimed to effectively capture the complex interplay between topics, emotions, and cognition in discussions - a capability traditional LDA models lack. Second, we incorporated Bayesian network principles to ensure result interpretability, crucial for educational stakeholders who need to understand the model’s reasoning. Third, we designed JECTM to handle MOOC discussion data’s characteristics, including its hierarchical structure (learner-post-word) and the co-occurrence of content and engagement features. These innovations enable JECTM to provide richer insights into discussions compared to conventional approaches.

To evaluate JECTM against baseline topic models, we employed normalized pointwise mutual information (NPMI) as our primary metric for topic coherence. NPMI quantitatively assesses how interpretable the topic-word distributions are from a human perspective (Hoyle et al., 2021). This metric has become the standard in recent literature, proving more effective than traditional measures like perplexity and entropy for evaluating topic model quality (Lau et al., 2014). We computed mean topic coherence across the top 10 words for each identified topic (Newman et al., 2010). We selected B-LDA (Qiu et al., 2013) and LDA (Blei et al., 2003) as baselines due to their widespread use in MOOC research (Gašević et al., 2018; Liu et al., 2019; Onan & Toçoğlu, 2020). B-LDA was specifically chosen for its similar structure in modeling behavioral patterns (such as response and like), while standard LDA served as a fundamental baseline that improved models should exceed.

Figure 4 displays the average NPMI scores of all models on our dataset, clearly comparing their performance. The results demonstrate that JCETM consistently outperformed both B-LDA and LDA across various topics. This superior performance can be attributed to JCETM’s incorporation of engagement information and ability to capture more nuanced relationships between words and topics. Furthermore, through a series of experiments, we determined that the optimal number of topics for JECTM on the dataset was 80.

The selection of 80 topics as the optimal number was based on comprehensive considerations across multiple dimensions. First, from a computational complexity perspective, this choice represents an optimal balance between topic granularity and processing efficiency. Experiments showed that when the number of topics exceeded 80, computational costs increased exponentially while model coherence improvements plateaued. Second, from a reliability perspective, the 80-topic model demonstrated significant stability with NPMI metrics reaching optimal levels across different topic models, indicating that this topic count effectively captures the fundamental discussion structure without overfitting. Third, from an interpretability perspective, 80 topics maintained both topic distinctiveness and reasonable costs for manual interpretation and coding. In conclusion, 80 topics not only achieved optimality in technical computational complexity metrics but also demonstrated good usability and interpretability in practical applications.

Coding Scheme of Topic Complexity

As illustrated in step 5 of Figure 2, we implemented a two-stage process for coding topic complexity. In the first stage, we classified topics into content-related and non-content-related topics following Wise et al. (2017). Content-related topics encompassed knowledge concepts directly from learning materials, while non-content-related topics covered course logistics like exam schedules. Two trained annotators independently classified the 80 JECTM-extracted topics using this scheme, achieving strong inter-rater reliability (Cohen’s kappa = 0.92, exceeding the 0.70 threshold) (Cohen, 1960). Any discrepancies were resolved by discussion until a consensus was reached between the two annotators.

For the second stage, we leveraged Bloom’s taxonomy of educational objectives to assess topic complexity levels (Bloom, 1956; Krathwohl, 2002; O’Riordan et al., 2021). This taxonomy represents a cumulative hierarchy where mastery of simpler categories precedes more complex ones (Krathwohl, 2002). As detailed in Table 3, our coding scheme incorporated three complexity levels aligned with six educational objectives. The annotators evaluated complexity based on key verbs and knowledge concept words, categorizing topics as: Low-Complexity Topics (LCT) focusing on remembering and understanding; Middle-Complexity Topics (MCT) emphasizing applying and analyzing; and High-Complexity Topics (HCT) involving evaluating and creating. For topics spanning multiple complexity levels, annotators assigned the level corresponding to the highest probability of key verbs in the topic-word distribution (Ertmer et al., 2011). Two annotators independently coded content-related topics’ complexity levels with strong reliability (Cohen’s kappa = 0.86, > 0.70), resolving any differences through negotiation to establish consensus.

Table 3.

Topic Complexity Coding Scheme based on Bloom’s Taxonomy of Educational Objectives (Bloom, 1956; Krathwohl, 2002; O’Riordan et al., 2021).

Topic Complexity	Objective	Description	Key Verb
Low-Complexity topic (LCT)	Remember	Requires learners to recall or recognize specific learned content through definition or recitation.	定义/Define, 描述/Describe, 命名/Label, 找出/Find, 列出/List,写出/ Write, and 说出/Tell.
	Understand	Requires learners to demonstrate comprehension by explaining or interpreting information.	解释/Explain, 重申/Restate, 分类Classify, 阐述/Translate, 改述/Paraphrase, and 诠释/Interpret.
Middle-Complexity topic (MCT)	Apply	Requires learners to transfer theoretical concepts to practical contexts.	解决/Solve, 应用/Apply, 使用/use, 改变/Change, 检查/Examine, 改进/Modify, and 实验/Experiment.
	Analyze	Requires learners to demonstrate analytical skills by comparing and contrasting relationships between different pieces of information.	分析/Analyze, 对比/Compare, 对照/Contrast, 推断/Infer, 区分/ Separate, and 局限/ Limit.
High-Complexity topic (HCT)	Evaluate	Requires learners to assess, judge, and argue the value of a decision or predict outcomes based on specific criteria.	评估/Evaluate, 决策/Decide, 选择/Choose, 推荐/Recommend, 评价/Assess, 优先级/Prioritize, and 批判/Critique.
	Create	Requires learners to synthesize information, evaluate the utility of different content elements, and produce novel arrangements or solutions.	创造/Create, 设计/Design, 计划/Plan, 整合/Combine, 形式化/Formulate, 创造 /Invent, and 开发/Develop.

Data Analysis

As depicted in step 6 of Figure 2, we employed a comprehensive analytical approach to address our research questions. For RQ1, we utilized BERT and JECTM to extract discussion topics, while topic-word distributions informed our coding of topic complexity levels. For RQ2, we conducted cross-tabulation analyses and chi-square tests to examine the associations between topic complexity and different forms of engagement, with standardized residuals providing deeper insights into these relationships. For RQ3, we leveraged Epistemic Network Analysis (ENA) to visualize and quantify the co-occurrence patterns of topic complexity and engagement types across high- and low-achieving groups (Shaffer et al., 2016). We complemented logistic regression analysis to investigate how topic complexity and engagement patterns jointly predict academic achievement (Soffer & Cohen, 2019).

Results

RQ1: what are the Discussion Topics and Their Levels of Complexity in the MOOC Forum?

Our analysis of discussion topics revealed several key findings regarding the distribution and complexity of content. Table 4 presents representative topics for each complexity category, showing their word probability distributions and distinguishing between terms from root posts (underlined) and response posts (non-underlined).

Table 4.

Complexity Levels and Their Representative Topics.

Content	Topic Complexity (Number of topics, %)	Prominent Topic (%)	Topic-Words Distribution
Non-content-related	Course logistics(11, 12.6%)	Topic-1(2.06%)	证书/Certificate	0.161	优秀/Excellence	0.004
			考试/Exam	0.156	学过/Learned	0.004
			参加/Participate	0.155	错过/Missed	0.004
			努力/Effort	0.039	期望/Expected	0.003
			时间/Time	0.011	收到/Received	0.003
			谢谢/Thank	0.007	提交/Submit	0.003
			测试/Test	0.005	表现/Performance	0.003
			加油/Cheer up	0.005	兴趣/Interesting	0.003
Content-related	LCTRemember and understand(17, 27.0%)	Topic-2(1.31%)	哪个/Which	0.101	精神分析/Psychoanalysis	0.011
			心理学/Psychology	0.058	作用/Impression	0.011
			理论/Theory	0.015	马斯洛/Maslow	0.005
			研究/Research	0.015	华生/Watson	0.009
			解释 / Explain	0.015	发展/Development	0.008
			人本主义/Humanistic	0.015	实验/Experiment	0.007
			弗洛伊德/Freudian	0.015	潜意识/Subconscious	0.006
			行为主义/Behaviorism	0.014	心理学家/Psychologists	0.005
	MCTApply and analysis(30, 30.9%)	Topic-3(1.72%)	课程/Courses	0.085	掌控/Mastering	0.008
			目标/Objectives	0.083	调节/Regulate	0.007
			案例/Examples	0.082	交际/Communication	0.005
			心理学/Psychology	0.080	心态/Mindset	0.005
			解决 / Solve	0.048	改变/Change	0.005
			帮助/Help	0.021	控制/Control	0.005
			情感/Emotion	0.018	能力/Ability	0.004
			行为/Behavior	0.014	问题/Problems	0.004
	HCTEvaluate and create(22, 29.5%)	Topic-4(2.07%)	父母/Parents	0.088	决定/Decisions	0.002
			影响/Influence	0.088	榜样/Role model	0.002
			儿童/Children	0.083	观点/Views	0.002
			原生/Origin	0.081	塑造/Shaping	0.002
			批判 / Critique	0.005	发展/Development	0.002
			养育/Parenting	0.003	系统/System	0.002
			教育/Education	0.003	个性/Personality	0.001
			成长/Growth	0.002	关系/Relationships	0.001

Note. Words underlined to indicate topic words of root posts. Words with underlining and italics were key verbs to identify topic complexity. Non-underlined words were the topic words of response posts.

LCT/MCT/HCT = Low/Middle/High-Complexity Topic.

Two key findings of our dataset support the robustness of our subsequent analyses. First, the predominance of content-related topics (n = 69, 87.4%) indicates that discussions were primarily focused on course material rather than irrelevant material, providing a solid foundation for analyzing content-related engagement patterns. Second, the balanced distribution across complexity levels (LCT = 27.0%, MCT = 30.9%, HCT = 29.5%) ensures sufficient data in each category for meaningful comparative analysis of how different complexity levels influence engagement.

The content-related topics exhibited distinct characteristics across complexity levels. Low-Complexity Topics (LCT, n = 17, 27.0%) centered on fundamental cognitive processes of remembering and understanding knowledge concepts. Topic-2 exemplifies this category, where learners explained psychological theories and received responses elaborating on various theoretical frameworks. Middle-Complexity Topics (MCT, n = 30, 30.9%) emphasized concept application and analysis, as demonstrated by Topic-3 where learners applied psychological principles to enhance social skills. High-Complexity Topics (HCT, n = 22, 29.5%) required sophisticated cognitive engagement through evaluation and creation. Topic-4 represents this highest complexity level, showing how learners critically evaluated the influence of original family dynamics on child development through multiple psychological lenses.

RQ2: How Does Topic Complexity Influence Emotional and Cognitive Engagement in MOOC Discussions?

RQ2a: How Does Topic Complexity Influence Emotional Engagement in MOOC Discussions?

Statistical analysis using the Chi-square test revealed a significant association between topic complexity and emotional engagement ( $χ_{(4)}^{2}$ = 1008.0, p < .001). As illustrated in Figure 5, examination of standardized residuals demonstrated that increasing topic complexity exhibited an inverse relationship with learners’ positive emotions while showing positive associations with confused and negative emotions. Specifically, LCT significantly enhanced positive emotions (85.80%, z = +15.20^*) while reducing both confusions (7.30%, z = −16.60^*) and negative emotions (6.90%, z = −12.70^*). In contrast, HCT was associated with decreased positive emotions (54.90%, z = −10.80^*) and elevated levels of both confusion (25.30%, z = +11.20^*) and negative emotions (19.80%, z = +9.90^*).

RQ2b: How Does Topic Complexity Influence Cognitive Engagement in MOOC Discussions?

Statistical analysis revealed a significant association between topic complexity and cognitive engagement ( $χ_{(4)}^{2}$ = 3573.37, p < .001). As shown in Figure 6, an examination of standardized residuals demonstrated that increasing topic complexity displayed positive associations with interactive and constructive cognition while showing an inverse relationship with active cognition. Specifically, LCT significantly diminished both interactive (4.00%, z = −20.90^*) and constructive cognition (46.10%, z = −12.60^*) while enhancing active cognition (49.90%, z = +45.00^*). MCT facilitated constructive cognition (65.30%, z = +65.30^*) while reducing active cognition (15.00%, z = −6.40^*). HCT promoted both interactive (26.60%, z = +14.10^*) and constructive cognition (68.50%, z = +6.10^*) while substantially decreasing active cognition (4.90%, z = −25.80^*).

RQ2c: How Does the Topic Complexity Jointly Influence Emotional and Cognitive Engagement?

Statistical analysis using the chi-squared test revealed a significant relationship between topic complexity and the co-occurrence patterns of emotional and cognitive engagement (

χ_{(16)}^{2}

= 4194.297, p < .001). As shown in Table 5, standardized residuals analysis demonstrated that different levels of topic complexity fostered distinct patterns of emotional and cognitive engagement co-occurrence. Specifically, LCT was associated with enhanced positive-constructive engagement (41.70%, z = +4.20^*) while reducing positive-active engagement (12.50%, z = −6.50^*). MCT showed a marked increase in positive-active engagement (44.90%, z = +45.50^*) paired with a significant decrease in positive-interactive engagement (3.00%, z = −14.90^*). HCT facilitated both positive-interactive (16.00%, z = +11.50^*) and confusion-constructive (17.20%, z = +12.40^*) engagement patterns, while substantially reducing positive-active engagement (3.10%, z = −26.00^*). These complex interrelationships highlight the importance of further research into the role of topic complexity, emotional engagement, and cognitive engagement in the prediction of academic achievement.

Table 5.

Analysis of Topic Complexity Triggering Emotional Engagement and Cognitive Engagement: Conditional Probability and Chi-Square Test Results.

Emotion-Cognition Triggered by Topic Complexity	Topic Complexity Level (%, z)
Emotion-Cognition Triggered by Topic Complexity	LCT	MCT	HCT
Positive-Interactive	11.10% (−0.70)	3.00% (−14.90^*)	16.00 % (+11.50^*)
Positive-Constructive	41.70% (+4.20^*)	37.90% (−0.80)	35.80% (−3.90^*)
Positive-Active	12.50% (−6.50^*)	44.90% (+45.50^*)	3.10% (−26.00^*)
Confusion-Interactive	6.60% (+3.30^*)	0.90% (−12.10^*)	7.20% (+5.20^*)
Confusion-Constructive	11.40% (−1.60)	3.50% (−14.80^*)	17.20% (+12.40^*)
Confusion-Active	1.90% (+0.80)	2.90% (+5.30^*)	1.00% (−4.70^*)
Negative-Interactive	2.00% (−1.00)	0.10% (−8.30^*)	3.40% (+7.10^*)
Negative-Constructive	12.20% (+0.50)	4.70% (−12.60^*)	15.50% (+8.60^*)
Negative-Active	0.70% (−2.70^*)	2.10% (+6.30^*)	0.80% (−1.70)

Note. LCT/MCT/HCT = Low/Middle/High-Complexity Topic. Standardized residual |z|^* > 1.96.

RQ3: How Do Topic Complexity, Emotional Engagement, and Cognitive Engagement Predict Academic Achievement in MOOC Forum Discussions?

Figure 7 illustrates the distinct patterns of interaction between topic complexity, emotional engagement, and cognitive engagement across high- and low-achievement learner groups. The mean epistemic networks for these groups are depicted in Figure 7(a) and (b) respectively. A subtraction network analysis (high minus low) presented in Figure 7(c) reveals compelling differences between the high-achievement group (depicted in blue, mean = −0.15, SD = 0.21) and low-achievement group (depicted in red, mean = 0.10, SD = 0.23) along the x-axis (α = 0.05, t (2577.57) = −29.56, p < .001, Cohen’s d = 1.11). The analysis reveals that the high-achievement group demonstrates stronger connections to HCT, maintains positive emotion, and engages in both constructive and interactive cognition. In contrast, the low-achievement group exhibits stronger associations with both HCT and MCT, displays more confused and negative emotion, and primarily engages in constructive cognition. Notably, when engaging with high-complexity topics, the high-achievement group demonstrates significantly higher levels of positive emotion and interactive cognition while experiencing markedly less confused and negative emotion compared to low-achievement learners.

Logistic regression analysis was conducted to examine the predictive relationships between topic complexity, emotional engagement, and cognitive engagement on academic achievement. Before analysis, multicollinearity diagnostics were performed to validate the regression model’s assumptions. The diagnostics confirmed that inter-indicator correlations remained below 0.90, indicating no problematic multicollinearity (Nistor et al., 2020). The analysis employed a stepwise approach to systematically eliminate non-significant predictors. The comprehensive results are presented in Table 6, where the odds ratio (OR), expressed as Exp(B), quantifies each variable’s predictive power relative to the baseline in determining high academic achievement.

Table 6.

Logistic Regression Results of Topic Complexity, Emotional Engagement, and Cognitive Engagement on Academic Achievement.

Model	R ²	Indicators	B	Exp (B)	SE	Wald	p
Model 1:Topic complexity(N = 2857)	0.280	HCT	0.219	1.245	0.025	78.01	0.000^***
		MCT	0.191	1.210	0.025	58.56	0.000^***
		LCT	0.177	1.193	0.031	32.03	0.000^***
Model 2:Topic complexity and emotion(N = 2857)	0.282	HCT-Positive	0.410	1.507	0.036	132.92	0.000^***
		MCT-Positive	0.217	1.242	0.031	48.21	0.000^***
		LCT-Positive	0.178	1.195	0.036	25.23	0.000^***
Model 3:Topic complexity and cognition(N = 2857)	0.294	HCT-Interactive	0.361	1.435	0.047	58.559	0.000^***
		HCT-Constructive	0.267	1.306	0.030	77.899	0.000^***
		LCT-Constructive	0.195	1.215	0.050	14.838	0.000^***
		MCT-Constructive	0.147	1.158	0.029	24.907	0.000^***
Model 4: Topic complexity, emotion, and cognition(N = 2857)	0.314	HCT-Positive-Interactive	0.561	1.752	0.065	75.479	0.000^***
		HCT-Positive-Constructive	0.480	1.616	0.044	117.119	0.000^***
		LCT-Positive-Constructive	0.211	1.235	0.060	12.358	0.000^***
		MCT-Positive-Constructive	0.201	1.222	0.039	26.404	0.000^***

Note. LCT/MCT/HCT = Low/Middle/High complexity topic. Nagelkerke R². Non-significant predictor variables were excluded by the stepwise method.

The analysis revealed several significant findings across four progressive models in Table 6. Model 1 (Topic Complexity, R² = 0.280) identified HCT as the strongest predictor of high achievement (OR = 1.245^***), followed by MCT (OR = 1.210^***) and LCT (OR = 1.193^***) in descending order. Model 2 (Topic Complexity and Emotion, R² = 0.282) demonstrated that HCT-positive (OR = 1.507^***) exhibited substantially stronger predictive capacity compared to MCT-positive (OR = 1.242^***) and LCT-positive (OR = 1.195^***). Model 3 (Topic Complexity and Cognition, R² = 0.293) showed that both HCT-interactive (OR = 1.435^***) and HCT-constructive (OR = 1.306^***) emerged as more powerful predictors than LCT-constructive (OR = 1.215^***) and MCT-constructive (OR = 1.158^***). Model 4 (Topic Complexity, Emotion and Cognition, R² = 0.314), the most comprehensive model, established that HCT-positive-interactive (OR = 1.752^***) and HCT-positive-constructive (OR = 1.616^***) engagement demonstrated markedly superior predictive strength compared to LCT-positive-constructive (OR = 1.235^***) and MCT-positive-constructive (OR = 1.222^***). These empirical findings reveal a consistent pattern: learners who engaged with high-complexity topics while maintaining positive emotions and demonstrating higher-order cognitive engagement (either interactive or constructive) showed significantly higher probabilities of achieving superior academic performance.

Discussions

This section discusses the key contributions and insights of our findings.

Methodological Advances in Analyzing MOOC Discussions at Scale

Our study makes several significant methodological contributions to the analysis of MOOC discussions at scale. First, we developed an innovative integrated two-step approach that combines automated topic modeling with engagement analysis, advancing beyond traditional methods that rely primarily on manual coding or basic text analysis techniques (Chen et al., 2020; Liu et al., 2022). This novel approach enables a rapid and comprehensive understanding of discussion topic trends, key question words, response content, and associated emotional and cognitive engagement patterns. The methodology provides a robust and efficient analytical framework for processing large-scale MOOC discussions.

Second, our research transcends the conventional single-dimensional focus on either cognitive or emotional engagement by integrating topic complexity analysis with both emotional and cognitive engagement measures (Liu et al., 2019, 2020). This multi-dimensional analytical approach reveals intricate patterns and relationships that would remain hidden in single-dimension analyses, offering deeper insights into the complex dynamics of online discussions.

Third, our study demonstrates the successful integration of advanced machine learning techniques with established educational theory in analyzing MOOC discussions. While previous researchers have employed algorithmic approaches to analyze online discussions (Onan & Toçoğlu, 2020; Wise & Cui, 2018), few have explicitly connected these technical methodologies to established theoretical frameworks of learning engagement. This synthesis of theoretical understanding and technological innovation enables a more nuanced and meaningful interpretation of discussion patterns and their implications.

Understanding the Interplay of Topic Complexity, Emotion, and Cognition

Our findings reveal several important relationships between topic complexity and learning engagement in MOOC discussions. While prior studies have acknowledged topic complexity’s role in influencing behavioral engagement (Lee & Recker, 2021), our research extends our understanding of its fine-grained effects on emotional and cognitive engagement. The results demonstrated that low-complexity topics generally elicited positive emotions but resulted in lower-order cognitive engagement. This may be because low-complexity topics are easier to understand and engage with, enhancing learners’ self-efficacy and positive emotions (Wigfield & Eccles, 2020).

In contrast, high-complexity topics promoted deeper cognitive processing while simultaneously increasing confusion and negative emotion. This aligns with previous research suggesting that high-complexity tasks place substantial demands on working memory resources, potentially leading to cognitive overload when the intrinsic cognitive load exceeds processing capacity (Merriënboer & Sweller, 2005). Insufficient cognitive resources may reduce learners’ sense of competence, generating negative emotions like confusion and anxiety (Pekrun & Linnenbrink-Garcia, 2012). Graesser (2020) further posit that confusion arises when learners encounter information conflicting with existing knowledge structures, necessitating deeper cognitive processing to resolve these cognitive conflicts. Therefore, this productive struggle may be necessary for developing higher-order thinking skills when appropriately scaffolded (Kapur, 2011). The key challenge lies in balancing cognitive challenge with emotional support to promote both deep thinking and positive emotional engagement.

The Role of Topic Complexity, Emotion, and Cognition on Academic Achievement

Our analysis reveals distinct engagement patterns between achievement groups. The discussion process of high-achieving learners demonstrated stronger associations between high-complexity topics, positive emotion, and both constructive and interactive cognitive engagement. The observed patterns align with past research suggesting that high-achieving learners typically demonstrate superior cognitive strategies and emotional regulation strategies, particularly in challenging learning tasks (Zimmerman & Schunk, 2011). These learners may be more adept at using peer interactions to construct knowledge, particularly when dealing with challenging content, and thus achieve higher academic achievement.

Conversely, the discussion process of low-achieving learners exhibited stronger correlations between middle-complexity topics, confusion, negative emotion, and primarily constructive cognition. A possible reason for this is that when encountering middle-complexity topics, these learners likely experience greater cognitive load. This increased cognitive burden leads them to express more confusion while promoting constructive cognition as they attempt to address the challenges (Graesser, 2020). The frustration experienced during the learning process often results in negative emotions (Merriënboer & Sweller, 2005). These findings suggest that emotion regulation plays an essential role in sustaining higher-order cognitive engagement (Dubovi, 2022). This underscores the importance of carefully designed discussion topics that provide sufficient cognitive challenge while fostering positive emotions to maintain sustained engagement.

Conclusions, Implications, and Future Works

In this study, a two-step approach was adopted to investigate the effect of topic complexity on emotional engagement, cognitive engagement, and academic achievement in MOOC discussions. This approach provides stakeholders with a quick and comprehensive overview of discussion topics and their emotional and cognitive engagement patterns. The evaluation showed that the approach outperformed known algorithms, which empowered efficient analysis of large-scale discussions.

This research provides practical implications for both instructors and learners. For instructors, our proposed two-step approach facilitates efficient monitoring of large-scale discussions, enabling instructors to track discussion topics and engagement patterns in real-time. We recommend three key implementation strategies: First, design topics progressively - begin with low-complexity topics to build confidence, then gradually introduce more complex topics while maintaining balanced emotional and cognitive engagement levels. Second, for challenging topics, provide targeted guidance, monitor emotional responses, and actively encourage peer interaction for collaborative problem-solving. Third, implement timely interventions by incorporating targeted guiding questions and prompts to stimulate meaningful discussion when topics become off-topic.

For learners, we propose several effective strategies to enhance their learning experience. First, our two-step approach serves as a learning analytics and self-regulation tool that generates discussion visualization reports (Chen et al., 2021; Ouyang et al., 2021), helping learners better understand their performance patterns and select appropriate discussion topics. Second, learners can track their engagement across topics of varying complexity, actively work to increase their cognitive engagement, and learn to regulate their emotions to maintain motivation. They can also enhance their learning strategies by observing and emulating the visualization reports of high-performing peers in complex discussions.

While our study makes significant contributions, several limitations warrant consideration. First, our analysis was confined to a single MOOC in psychology, future research should examine these relationships across diverse subject areas and learning contexts. Second, subsequent studies could investigate additional influential factors such as learner characteristics, prior knowledge levels, and motivational aspects that may moderate the relationships we observed. Third, our innovative two-step approach shows promise for application beyond MOOCs to other educational contexts, including classroom dialogues and agent-based conversations (e.g., AI tutors and chatbots). Fourth, while we have adopted achievement measurement methods commonly used in existing studies (Doo & Kim, 2024; Liu et al., 2022; Liu et al., 2024), future research could focus on implementing discussion-specific measures that can more accurately capture the unique characteristics of discussions. Fifth, the current three-dimensional emotional engagement framework could be expanded to include a broader range of emotional categories for more comprehensive measurement.

Footnotes

Author Contributions

Mr Shiqi Liu: Writing - Original Draft, Writing - Review & Editing, Methodology

Dr Sannyuya Liu: Writing - Original Draft, Conceptualization, Project administration

Dr Xian Peng: Writing - Review & Editing, Methodology, Writing - Original Draft

Dr Jianwen Sun: Writing - Original Draft, Writing - Review & Editing, Resources

Dr Zhi Liu: Writing - Original Draft, Writing - Review & Editing, Resources.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was financially supported by the National Science and Technology Major Project (2022ZD0117101), National Natural Science Foundation of China (62377016, 62293554, 62293550, 62407023), Hubei Provincial Natural Science Foundation of China (2023AFA020), the 2024 “AI + Examination and Evaluation” Special Teaching Reform Project of Central China Normal University (CCNU24JG10), and the Major Project of Philosophy and Social Science Research in Colleges and Universities of Jiangsu Province (2024SJZD074).

ORCID iD

Shiqi Liu

Data Availability Statement

Data is available on request from the authors.*

Appendix

Equation (A1) denoted the joint probability equation of JECTM. (A1)

\begin{array}{l} p (Z, W, C, E ∣ μ, η, β, α) \\ = p (W ∣ Z, β) p (C ∣ Z, η) p (E ∣ Z, μ) p (Z ∣ α) \\ = \prod_{t = 1}^{T} \frac{Δ (n_{t}^{w} + β)}{Δ (β)} \cdot \prod_{t = 1}^{T} \frac{Δ (n_{t}^{c} + η)}{Δ (η)} \cdot \prod_{t = 1}^{T} \frac{Δ (n_{t}^{e} + μ)}{Δ (μ)} \cdot \prod_{l = 1}^{L} \frac{Δ (n_{l}^{t} + α)}{Δ (α)} \\ , Δ (\vec{\cdot}) = \frac{\prod_{i = 1}^{k} Γ (\cdot_{i})}{Γ (\sum_{i = 1}^{k} \cdot_{i})} \end{array}

Equation (A2) presented the derivation process of the Gibbs sampling formula. (A2)

\begin{array}{l} p (z_{d} ∣ Z_{\neg_{d}}, W, E, C) = \propto \frac{p (Z, W, E, C ∣ η, μ, β, α)}{p (Z_{\neg_{d}}, W, E_{\neg_{d}}, C_{\neg_{d}} ∣ η, μ, β, α)} \\ = \prod_{t = 1}^{T} \frac{Δ (n_{t}^{w} + β)}{Δ (n_{t, \neg d}^{w} + β)} \prod_{t = 1}^{T} \frac{Δ (n_{t}^{c} + η)}{Δ (n_{t, \neg_{d}}^{c} + η)} \cdot \prod_{t = 1}^{T} \frac{Δ (n_{t}^{e} + μ)}{Δ (n_{t, \neg_{d}}^{e} + μ)} \cdot \prod_{l = 1}^{L} \frac{Δ (n_{l}^{t} + α)}{Δ (n_{l, \neg_{d}}^{t} + α)} \\ = \frac{\prod_{w = 1}^{V} \prod_{i = 1}^{n^{w}} (n_{z, \neg_{d}}^{w} + β + i - 1)}{\prod_{j = 1}^{n^{W}} (\sum_{w = 1}^{V} n_{z, \neg_{d}}^{w} + V β + j - 1)} \cdot \frac{n_{z, \neg_{d}}^{c_{d}} + η}{\sum_{c = 1}^{C} n_{z, \neg_{d}}^{c} + C η} \cdot \frac{n_{z, \neg_{d}}^{e_{d}} + μ}{\sum_{s = 1}^{E} n_{z, \neg_{d}}^{e} + E μ} \cdot \frac{n_{l, \neg_{d}}^{z} + α}{\sum_{t = 1}^{T} n_{l, \neg_{d}}^{t} + T α} \\ Δ (\vec{\cdot}) = \frac{\prod_{i = 1}^{k} Γ (\cdot_{i})}{Γ (\sum_{i = 1}^{k} \cdot_{i})} \end{array}

The model parameters - learner-topics distribution

θ_{l}

, topic-words distribution

ϕ_{t}

, topic-emotions distribution

π_{t e}

, and topic-cognitions distribution

ψ_{t c}

are calculated using Equations (A3–A6). Specifically, in equation (A3),

n_{l}^{t}

denotes the frequency of posts assigned to the topic

z_{t}

for a given learner l. In equation (A4),

n_{t}^{w}

denotes the frequency of word w occurring in topic t. In equation (A5),

n_{t}^{e}

denotes the frequency of posts labeled with emotion e under topic t. In equation (A6),

n_{t}^{c}

denotes the frequency of posts labeled with cognitive label c under topic t. Through empirical optimization, the prior parameters α, β,

μ

, and

η

were set as 1, 0.01, 0.1, and 0.1, respectively. (A3)

θ_{l} = \frac{n_{l}^{t} + α}{\sum_{t = 1}^{T} n_{l}^{t} + T α}

(A4)

ϕ_{t, w} = \frac{n_{t}^{w} + β}{\sum_{w = 1}^{V} n_{t}^{w} + V β}

(A5)

π_{t, e} = \frac{n_{t}^{e} + μ}{\sum_{s = 1}^{E} n_{t}^{e} + E μ}

(A6)

ψ_{t, c} = \frac{n_{t}^{c} + η}{\sum_{c = 1}^{C} n_{t}^{c} + C η}

We employed Gibbs sampling for JECTM inference, with the detailed process outlined in Algorithm A1. Before analysis, we preprocessed the Chinese discussion posts through several steps: First, following LDA modeling assumptions, we combined root posts with their corresponding response posts to optimize topic-word distribution coherence and interpretability. Second, we utilized HanLP for word tokenization. Third, we created a domain-specific dictionary to ensure accurate word segmentation, incorporating key terms like “内在动机/intrinsic motivation” and “发展心理学/developmental psychology”. Fourth, we cleaned the text by removing stop words (e.g., “和/and,” “的/of”), low-frequency words (occurring ≤ 5 times), and noise elements (URLs and punctuation).

Algorithm A1: Gibbs sampling for JECTM
Import: Pre-processed learner discussion posts. The iteration times of Gibbs sampling. topics number: T hyper-parameter: $α, β, η, μ$ Output: Learner-topics distribution $θ_{l}$ , topic-words distribution $ϕ_{t}$ Topic-cognitions distribution $ψ_{t c}$ , topic-emotions distribution $π_{t e}$ Procedure Gibbs sampling: 1: Begin procedure 2: for each learner l = 1,…,l do 3: for l’s d-th post, d = 1,…, D do 4: for each word w = 1, …, W do 5: Randomly assign a topic to $z_{l, d}$ 6: end for 7: end for 8: end for 9: for each Gibbs Sampling iteration, do 10: for each learner l = 1,…, L do 11: for l’s d-th document, d = 1,…, D do 12: Gibbs sample topic $z_{l, d}$ according to equation (A2) 13: end for 14: end for 15: end for 16: end for 17: Estimate model parameters $θ_{l}$ , $ϕ_{t}$ , $π_{t e}$ , $ψ_{t c}$ , according to equations (A3–A6) 18: End procedure

Algorithm A1: Gibbs sampling for JECTM

Import:
Pre-processed learner discussion posts.
The iteration times of Gibbs sampling.
topics number: T
hyper-parameter:

α, β, η, μ

Output:
Learner-topics distribution

θ_{l}

, topic-words distribution

ϕ_{t}

Topic-cognitions distribution

ψ_{t c}

, topic-emotions distribution

π_{t e}

Procedure Gibbs sampling:
1: Begin procedure
2: for each learner l = 1,…,l do
3: for l’s d-th post, d = 1,…, D do
4: for each word w = 1, …, W do
5: Randomly assign a topic to

z_{l, d}

6: end for
7: end for
8: end for
9: for each Gibbs Sampling iteration, do
10: for each learner l = 1,…, L do
11: for l’s d-th document, d = 1,…, D do
12: Gibbs sample topic

z_{l, d}

according to equation (A2)
13: end for
14: end for
15: end for
16: end for
17: Estimate model parameters

θ_{l}

ϕ_{t}

π_{t e}

ψ_{t c}

, according to equations (A3–A6)
18: End procedure

Author Biographies

Shiqi Liu is a PhD student currently working on a Ph.D. degree in Educational Technology at the Faculty of Artificial Intelligence in Education, Central China Normal University (CCNU). He is a visiting Ph.D. student at the Centre for Learning Analytics, Faculty of Information Technology, Monash University. His research interests include natural language processing and learning analytics.

Sannyuya Liu is currently a Professor, Ph.D. Supervisor, and Director at the Faculty of Artificial Intelligence in Education, CCNU. His research interests include computer applications, artificial intelligence, and educational information technology.

Xian Peng is currently an Associate Professor with the Faculty of Artificial Intelligence in Education and National Engineering Research Center of Educational Big Data, CCNU. His research interests include learning analytics and educational data mining.

Jianwen Sun is currently a Professor and PhD Supervisor with the Faculty of Artificial Intelligence in Education and National Engineering Research Center of Educational Big Data, CCNU. His research interests include educational data mining, learning analytics, and intelligent tutoring systems.

Zhi Liu is a Professor and PhD supervisor at the National Engineering Research Center of Education Big Data, Faculty of Artificial Intelligence in Education, CCNU. He also holds a position as a guest researcher at the Computer Science Institute, Humboldt University of Berlin. His research interests include educational data mining, learning analytics, and intelligent tutoring systems.

References

Atapattu

Thilakaratne

Vivian

Falkner

(2019). Detecting cognitive engagement using word embeddings within an online teacher professional development community. Computers & Education, 140(20), Article 103594. https://doi.org/10.1016/j.compedu.2019.05.020

Birjali

Kasri

Beni-Hssane

(2021). A comprehensive survey on sentiment analysis: Approaches, challenges and trends. Knowledge-Based Systems, 226(8), Article 107134. https://doi.org/10.1016/j.knosys.2021.107134

Blei

D. M.

A. Y.

Jordan

M. I.

(2003). Latent Dirichlet allocation. The Journal of Machine Learning Research, 3(2), 993–1022.

Bloom

B. S.

(1956). Taxonomy of educational objectives: The classification of educational goals, 1st ed. Longman Group.

Buhr

E. E.

Daniels

L. M.

Goegan

L. D.

(2019). Cognitive appraisals mediate relationships between two basic psychological needs and emotions in a massive open online course. Computers in Human Behavior, 96(9), 85–94. https://doi.org/10.1016/j.chb.2019.02.009

Chen

C.-M.

M.-C.

Chang

W.-C.

Chen

X.-X.

(2021). Developing a topic analysis instant feedback system to facilitate asynchronous online discussion effectiveness. Computers & Education, 163(4), Article 104095. https://doi.org/10.1016/j.compedu.2020.104095

Chen

C. K.

(2020). Sustaining online academic discussions: Identifying the characteristics of messages that receive responses. Computers & Education, 156(1), Article 103938. https://doi.org/10.1016/j.compedu.2020.103938

Chevrier

Muis

K. R.

Trevors

G. J.

Pekrun

Sinatra

G. M.

(2019). Exploring the antecedents and consequences of epistemic emotions. Learning and Instruction, 63(6), Article 101209. https://doi.org/10.1016/j.learninstruc.2019.05.006

Chi

M. T. H.

Wylie

(2014). The ICAP framework: Linking cognitive engagement to active learning outcomes. Educational Psychologist, 49(4), 219–243. https://doi.org/10.1080/00461520.2014.965823

10.

Cohen

(1960). A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1), 37–46. https://doi.org/10.1177/001316446002000104

11.

Dalipi

Zdravkova

Ahlgren

(2021). Sentiment analysis of students’ feedback in MOOCs: A systematic literature review. Frontiers in Artificial Intelligence, 4(1), Article 728708. https://doi.org/10.3389/frai.2021.728708

12.

Demir

Ö.

Cinar

Keskin

(2023). Participation style and social anxiety as predictors of active participation in asynchronous discussion forums and academic achievement. Education and Information Technologies, 28(9), 1–22. https://doi.org/10.1007/s10639-022-11517-3

13.

Devlin

Chang

M.-W.

Lee

Toutanova

(2019). Bert: Pre-Training of deep bidirectional Transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), Minneapolis USA, June 2-7, pp. 4171–4186.

14.

Dhawal

Archisha

(2023). 202K courses, 662M enrollments: Breaking down udemy’s massive catalog. https://www.classcentral.com/report/udemy-by-the-numbers/

15.

Dillon

Bosch

Chetlur

Wanigasekara

Ambrose

G. A.

Sengupta

D’Mello

S. K.

(2016). Student emotion, Co-occurrence, and dropout in a MOOC context. International Educational Data Mining Society . https://eric.ed.gov/?id=ED592723

16.

Ding

(2019). Applying gamifications to asynchronous online discussions: A mixed methods study. Computers in Human Behavior, 91(9), 1–11. https://doi.org/10.1016/j.chb.2018.09.022

17.

D’Mello

Dieterle

Duckworth

(2017). Advanced, analytic, automated (AAA) measurement of engagement during learning. Educational Psychologist, 52(2), 104–123. https://doi.org/10.1080/00461520.2017.1281747

18.

D’Mello

Lehman

Pekrun

Graesser

(2014). Confusion can be beneficial for learning. Learning and Instruction, 29(3), 153–170. https://doi.org/10.1016/j.learninstruc.2012.05.003

19.

Doo

M. Y.

Kim

(2024). The relationship between learning engagement and learning outcomes in online learning in higher education: A meta-analysis study. Distance Education, 45(1), 60–82. https://doi.org/10.1080/01587919.2024.2303484

20.

Dubovi

(2022). Cognitive and emotional engagement while learning with VR: The perspective of multimodal methodology. Computers & Education, 183(4), Article 104495. https://doi.org/10.1016/j.compedu.2022.104495

21.

Ertmer

P. A.

Sadaf

Ertmer

D. J.

(2011). Student-content interactions in online courses: The role of question prompts in facilitating higher-level engagement with course content. Journal of Computing in Higher Education, 23(2–3), 157–186. https://doi.org/10.1007/s12528-011-9047-6

22.

Gašević

Joksimović

Eagan

B. R.

Shaffer

D. W.

(2018). SENS: Network analytics to combine social and cognitive perspectives of collaborative learning. Computers in Human Behavior, 92(3), 562–577. https://doi.org/10.1016/j.chb.2018.07.003

23.

Gašević

Kovanović

Joksimović

(2017). Piecing the learning analytics puzzle: A consolidated model of a field of research and practice. Learning: Research and Practice, 3(1), 63–78. https://doi.org/10.1080/23735082.2017.1286142

24.

Geng

Niu

Feng

Huang

(2020). Understanding the focal points and sentiment of learners in MOOC reviews: A machine learning and SC-LIWC-based approach. British Journal of Educational Technology, 51(5), 1785–1803. https://doi.org/10.1111/bjet.12999

25.

Graesser

A. C.

(2020). Emotions are the experiential glue of learning environments in the 21st century. Learning and Instruction, 70(9), Article 101212. https://doi.org/10.1016/j.learninstruc.2019.05.009

26.

Han

Z.-M.

Huang

C.-Q.

J.-H.

Tsai

C.-C.

(2021). Identifying patterns of epistemic emotions with respect to interactions in massive online open courses using deep learning and social network analysis. Computers in Human Behavior, 122(18), Article 106843. https://doi.org/10.1016/j.chb.2021.106843

27.

(2021). HanLP: Han language processing. [Python]. https://github.com/hankcs/HanLP

28.

Hou

Zhu

Zheng

Zhang

Huang

Zhong

Ker

C. L.

(2024). Prompt-based and fine-tuned GPT models for context-dependent and -independent deductive coding in social annotation. In Proceedings of the 14th Learning Analytics and Knowledge Conference, Tokyo Japan, March 18-22, pp. 518–528. https://doi.org/10.1145/3636555.3636910

29.

Hoyle

Goel

Hian-Cheong

Peskov

Boyd-Graber

Resnik

(2021). Is automated topic model evaluation broken? The incoherence of coherence. Advances in Neural Information Processing Systems, 34(26), 2018–2033. https://proceedings.neurips.cc/paper/2021/hash/0f83556a305d789b1d71815e8ea4f4b0-Abstract.html.

30.

Iqbal

Rakovic

Chen

Ferreira Mello

Fan

Fiorentino

Radi Aljohani

Gasevic

(2023). Towards automated analysis of rhetorical categories in students essay writings using Bloom’s taxonomy. In LAK23: 13th International Learning Analytics and Knowledge Conference, Austin USA, March 13-17, pp. 418–429. https://doi.org/10.1145/3576050.3576112

31.

Kapur

(2011). Temporality matters: Advancing a method for analyzing problem-solving processes in a computer-supported collaborative environment. International Journal of Computer-Supported Collaborative Learning, 6(1), 39–56. https://doi.org/10.1007/s11412-011-9109-9

32.

Kong

Liu

Chen

Liu

Tang

(2025). Exploratory study of an AI-supported discussion representational tool for online collaborative learning in a Chinese university. The Internet and Higher Education, 64(2), Article 100973. https://doi.org/10.1016/j.iheduc.2024.100973

33.

Krathwohl

D. R.

(2002). A revision of Bloom’s taxonomy: An overview. Theory Into Practice, 41(4), 212–218. https://doi.org/10.1207/s15430421tip4104_2

34.

Lau

J. H.

Newman

Baldwin

(2014). Machine reading tea leaves: Automatically evaluating topic coherence and topic model quality. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, Gothenburg Sweden, April 26-30, pp. 530–539. https://doi.org/10.3115/v1/E14-1056

35.

Lave

Wenger

(1991). Situated Learning: Legitimate Peripheral participation. Cambridge University Press.

36.

Lee

J.-E.

Recker

(2021). The effects of instructors’ use of online discussions strategies on student participation and performance in university online introductory mathematics courses. Computers & Education, 162(5), Article 104084. https://doi.org/10.1016/j.compedu.2020.104084

37.

Liu

Peng

Yang

(2022). Automated detection of emotional and cognitive engagement in MOOC discussions to predict learning achievement. Computers & Education, 181(9), Article 104461. https://doi.org/10.1016/j.compedu.2022.104461

38.

Liu

Sha

Zeng

Gasevic

Liu

(2024). Annotation Guidelines-based knowledge Augmentation: Towards enhancing large language models for educational text classification No. arXiv:2406.00954. arXiv. https://doi.org/10.48550/arXiv.2406.00954

39.

Liu

Zhang

Liu

(2020). Investigating the relationship between learners’ cognitive participation and learning outcome in asynchronous online discussion forums. In Proceedings of the 12th International Conference on Computer Supported Education, Prague Czech Republic, May 2-4, pp. 26–33.

40.

Liu

Tang

Ouyang

Long

Liu

(2024). Profiling students’ learning engagement in MOOC discussions to identify learning achievement: An automated configurational approach. Computers & Education, 219(9), Article 105109. https://doi.org/10.1016/j.compedu.2024.105109

41.

Liu

Yang

Rüdian

Liu

Zhao

Wang

(2019). Temporal emotion-aspect modeling for discovering what students are concerned about in online course forums. Interactive Learning Environments, 27(5–6), 598–627. https://doi.org/10.1080/10494820.2019.1610449

42.

Malinen

(2015). Understanding user participation in online communities: A systematic literature review of empirical studies. Computers in Human Behavior, 46(4), 228–238. https://doi.org/10.1016/j.chb.2015.01.004

43.

Merriënboer

J. J. G. V.

Sweller

(2005). Cognitive load theory and complex learning: Recent developments & future directions. Educational Psychology Review, 17(5), Article 147177.

44.

Mohammed

Omar

(2020). Question classification based on Bloom’s taxonomy cognitive domain using modified TF-IDF and word2vec. PLoS One, 15(3), Article e0230442. https://doi.org/10.1371/journal.pone.0230442

45.

Newman

Lau

J. H.

Grieser

Baldwin

(2010). Automatic evaluation of topic coherence. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Los Angeles USA, June 1–6, pp. 100–108. https://www.aclweb.org/anthology/N10-1012

46.

Nistor

Dascalu

Tarnai

Trausan-Matu

(2020). Predicting newcomer integration in online learning communities: Automated dialog assessment in blogger communities. Computers in Human Behavior, 105(12), Article 106202. https://doi.org/10.1016/j.chb.2019.106202

47.

Onan

Toçoğlu

(2020). Weighted word embeddings and clustering‐based identification of question topics in MOOC discussion forum posts. Computer Applications in Engineering Education, 29(4), 675–689. https://doi.org/10.1002/cae.22252

48.

O’Riordan

Millard

D. E.

Schulz

(2021). Is critical thinking happening? Testing content analysis schemes applied to MOOC discussion forums. Computer Applications in Engineering Education, 29(4), 690–709. https://doi.org/10.1002/cae.22314

49.

Orji

F. A.

Vassileva

(2023). Predicting the persuasiveness of influence strategies from student online learning behaviour using machine learning methods. Journal of Educational Computing Research, 61(7), 1410–1429. https://doi.org/10.1177/07356331231178873

50.

Ouyang

(2021). Using three social network analysis approaches to understand computer-supported collaborative learning. Journal of Educational Computing Research, 59(7), 1401–1424. https://doi.org/10.1177/0735633121996477

51.

Ouyang

Chen

(2021). Effect of three network visualizations on students’ social-cognitive engagement in online discussions. British Journal of Educational Technology, 52(6), 2242–2262. https://doi.org/10.1111/bjet.13126

52.

Pan

Zhou

Feng

(2022). Are students happier the more they learn? – Research on the influence of course progress on academic emotion in online learning. Interactive Learning Environments, 31(10), 1–21. https://doi.org/10.1080/10494820.2022.2052110

53.

Pekrun

Linnenbrink-Garcia

(2012). Academic emotions and student engagement. In Christenson

S. L.

Reschly

A. L.

Wylie

(Eds.), Handbook of research on student engagement (pp. 259–282). Springer US. https://doi.org/10.1007/978-1-4614-2018-7_12

54.

Peng

Han

Ouyang

Liu

(2020). Topic tracking model for analyzing student-generated posts in SPOC discussion forums. International Journal of Educational Technology in Higher Education, 17(1), 35. https://doi.org/10.1186/s41239-020-00211-4

55.

Peng

(2020). Investigating learners’ behaviors and discourse content in MOOC course reviews. Computers & Education, 143, Article 103673. https://doi.org/10.1016/j.compedu.2019.103673

56.

Qiu

Zhu

Jiang

(2013). It is not just what we say, but how we say them: LDA-based behavior-topic model. In Proceedings of the 2013 SIAM International Conference on Data Mining, Austin USA, May 2-4, pp. 794–802. https://doi.org/10.1137/1.9781611972832.88

57.

Ramirez-Arellano

Bory-Reyes

Hernández-Simón

L. M.

(2019). Emotions, motivation, cognitive–metacognitive strategies, and behavior as predictors of learning performance in blended learning. Journal of Educational Computing Research, 57(2), 491–512. https://doi.org/10.1177/0735633117753935

58.

Rivera

D. A.

Frenay

Paquot

(2024). The role of MOOC forum discussion tasks in learners’ cognitive engagement. Journal of Computer Assisted Learning, 40(5), 2103, 2120. https://doi.org/10.1111/jcal.13001

59.

Shaffer

D. W.

Collier

Ruis

A. R.

(2016). A tutorial on epistemic network analysis: Analyzing the structure of connections in cognitive, social, and interaction data. Journal of Learning Analytics, 3(3), Article 3. https://doi.org/10.18608/jla.2016.33.3

60.

Shaikh

Daudpotta

S. M.

Imran

A. S.

(2021). Bloom’s learning outcomes’ automatic classification using LSTM and pretrained word embeddings. IEEE Access, 9(6), 117887–117909. https://doi.org/10.1109/ACCESS.2021.3106443

61.

Soffer

Cohen

(2019). Students’ engagement characteristics predict success and completion of online courses. Journal of Computer Assisted Learning, 35(3), 378–389. https://doi.org/10.1111/jcal.12340

62.

Sun

J. C.-Y.

Rueda

(2012). Situational interest, computer self-efficacy and self-regulation: Their impact on student engagement in distance education. British Journal of Educational Technology, 43(2), 191–204. https://doi.org/10.1111/j.1467-8535.2010.01157.x

63.

Tang

Xing

Pei

(2019). Time really matters: Understanding the temporal dimension of online learning using educational data mining. Journal of Educational Computing Research, 57(5), 1326–1347. https://doi.org/10.1177/0735633118784705

64.

Vygotsky

L. S.

(1978). Mind in society. Harvard University Press. https://book.douban.com/subject/2266607/

65.

Wang

Shan

Zheng

Guo

Chen

(2023). Can chatgpt detect student talk moves in classroom discourse? A preliminary comparison with bert. In Proceedings of the 16th International Conference on Educational Data Mining, Bengaluru India, July 11-14, pp. 515–519. https://doi.org/10.5281/zenodo.8115772

66.

Wang

Wen

Rosé

C. P.

(2016). Towards triggering higher-order thinking behaviors in MOOCs. In Proceedings of the Sixth International Conference on Learning Analytics & Knowledge, Edinburgh Scotland, April 25-29, pp. 398–407. https://doi.org/10.1145/2883851.2883964

67.

Wang

Yang

Wen

Koedinger

Rosé

C. P.

(2015). Investigating how student’s cognitive behavior in MOOC discussion forums affect learning gains. International Educational Data Mining Society .

68.

Wei

Saab

Admiraal

(2021). Assessment of cognitive, behavioral, and affective learning outcomes in massive open online courses: A systematic literature review. Computers & Education, 163(7), Article 104097. https://doi.org/10.1016/j.compedu.2020.104097

69.

Wigfield

Eccles

J. S.

(2020). 35 years of research on students' subjective task values and motivation: A look back and a look forward. In Elliot

A. J.

(Ed.), Advances in motivation science (7, pp. 161–198). Elsevier. https://doi.org/10.1016/bs.adms.2019.05.002

70.

Wise

A. F.

Cui

(2018). Learning communities in the crowd: Characteristics of content related interactions and social relationships in MOOC discussion forums. Computers & Education, 122(10), 221–242. https://doi.org/10.1016/j.compedu.2018.03.021

71.

Wise

A. F.

Cui

Jin

Vytasek

(2017). Mining for gold: Identifying content-related MOOC discussion threads across domains through linguistic modeling. The Internet and Higher Education, 32(1), 11–28. https://doi.org/10.1016/j.iheduc.2016.08.001

72.

Wise

A. F.

Knight

Shum

S. B.

(2021). Collaborative learning analytics. In Cress

Rosé

Wise

A. F.

Oshima

(Eds.), International handbook of computer-supported collaborative learning (1st ed., pp. 425–443). Springer International Publishing. https://doi.org/10.1007/978-3-030-65291-3_23

73.

Xing

(2023). Mining large open online learning networks: Exploring community dynamics and communities of performance. Journal of Educational Computing Research, 61(2), 390–415. https://doi.org/10.1177/07356331221113613

74.

Yang

Tang

Hao

Rose

J. R.

(2022). Untangling chaos in discussion forums: A temporal analysis of topic-relevant forum posts in MOOCs. Computers & Education, 178, Article 104402. https://doi.org/10.1016/j.compedu.2021.104402

75.

Yao

Zhang

Wei

Jin

Zhang

Chen

(2017). Incorporating knowledge graph embeddings into topic modeling. In Thirty-First AAAI Conference on Artificial Intelligence, San Francisco USA, February 4-9. https://doi.org/10.1609/aaai.v31i1.10951

76.

Zimmerman

B. J.

Schunk

D. H.

(2011). Self-regulated learning and performance: An introduction and an overview. In Handbook of self-regulation of learning and performance. Routledge.

77.

Zou

Pan

Cai

Liu

(2021). Exploring the relationship between social presence and learners’ prestige in MOOC discussion forums using automated content analysis and social network analysis. Computers in Human Behavior, 115(11), 106582. https://doi.org/10.ghmph7