Project-based learning plays a crucial role in computing education. However, its open-ended nature makes tracking project development and assessing success challenging. We investigate how dialogue and system interaction logs predict project quality during collaborative, project-based AI learning of 94 middle school students working in pairs. We used linguistic features from dialogue transcripts and behavioral features from system logs to predict three project quality outcomes: productivity (number of training phrases), content richness (word density), and lexical variation (word diversity) of chatbot training phrases. We compared the predictive accuracy of each modality and a fusion of the modalities. Results indicate log data better predicts productivity, while dialogue data is more efective for content richness. Both modalities modestly predict lexical variation. Multimodal fusion improved predictions for productivity and lexical variation of training phrases but not content richness. These findings suggest that the value of multimodal fusion depends on the specific learning outcome. The study contributes to multimodal learning analytics by demonstrating the nuanced interplay between behavioral and linguistic data in assessing student learning progress in open-ended AI learning environments.
Project-based learning (PBL) is crucial in STEM, especially in computing, where students collaboratively
create artifacts [
CEUR
ceur-ws.org
Researchers have explored various data sources and analytical approaches to better understand and support student learning during PBL. A longstanding tradition in learning analytics involves inferring cognitive and metacognitive processes from log and language data [18]. While researchers have leveraged log interactions to model learning trajectories [17, 19], and language data to infer self-regulated learning strategies [20], these modalities are rarely used for joint prediction. Integrating dialogue-based and log-based features may ofer a more holistic view of student collaborative learning analytics [21, 22]. While prior research has focused on predicting learning gains using multimodal collaboration features [23], less attention has been given to the temporal dynamics of learning [18]. We address this gap by predicting the quality of individual learning events in collaborative learning, that is, the quality of training phrases while students program a chatbot. Specifically, we compare the predictive capabilities of dialogue and log data in assessing the quality of student projects in an AI learning context where students collaboratively develop chatbots. We examine whether combining diferent modalities of data enhances the predictive accuracy of project quality compared to using them independently:
RQ1: How well can student project quality be predicted from single modalities (dialogue, log data)? RQ2: To what extent does the multimodal fusion of these data sources enhance predictive accuracy?
Multimodal approaches to mining data and modeling are increasingly common in educational data mining (EDM) [24, 25]. A recent review by Chango et al. [26] traces emerging trends in the use of multimodal in typical EDM contexts, including online learning environments, classroom studies, and blended learning contexts. Data collected in these sites can range from text data to log data, physiological learner data such as eye-tracking, and spatiotemporal data such as teacher positions [25].
Natural language data, especially, is increasingly used for foundational EDM prediction tasks. For example, Zhang et al. [20] utilized large language model (LLM) embeddings to predict self-regulated learning stages from think-aloud protocols, illustrating the potential of text-based data in understanding cognitive processes. Scarlatos et al. [27] explored using LLM for knowledge tracing from learners’ dialogue. Borchers et al. [19] fused data from peer tutoring chat collaborations with tutoring system problem-solving logs to model learning rates in relation to diferent dialogue acts. However, these approaches have rarely combined language data with log data to jointly model learning.
The present study builds on this body of work by applying text mining techniques based on automated transcriptions and embedding [20] with log data for a novel prediction task: predicting learner project quality in AI education. Specifically, we incorporate transcriptions of students’ collaboration as a data source in EDM, combining them with established log data to enhance predictive accuracy. By testing the combined predictive power of log data and transcription-derived features, we test whether multimodal fusion improves the prediction of learning processes during collaboration. We contribute to the ongoing discourse on the utility of multimodal data in enhancing educational predictions and insights.
Collaborative learning systems ofer opportunities to mine multimodal data from learning-system and learner-learner interactions [28, 29].1 These systems provide insights into learning patterns through modeling tutoring rates and collaborative behaviors [19], which have been used for feedback tools [31, 22]. However, selecting meaningful outcomes and learning constructs for multimodal analysis remains challenging. The present study embraces this challenge in a novel predictive context, where students learn to program chatbots using natural language inputs. We predict the quality of 1We define collaboration as the coordinated efort to jointly engage in problem-solving and knowledge construction [ 30]. student-created training phrases during learning. Insights from this research have implications for other collaborative systems and performance prediction from collaborative characteristics [19, 23, 32, 33].
In addition to peer collaboration interactions, research has demonstrated how dialogue can be leveraged to support learner-system interactions. For instance, intelligent agents can adapt instruction based on detected learning patterns [34]. In contrast, teachable agents provide practice problems and assessments [35] and build rapport with learners through lexical adaptation toward learners [36]. Similarly, adaptive dialogue systems such as EER-tutor have demonstrated improved learning gains by considering prior student errors [37]. Finally, Gaze Tutor [38] leveraged gaze data to detect disengagement and adapt dialogue to re-engage students. The present study builds on this foundation by seeking to develop meaningful predictors of student learning performance during collaboration, making progress toward detectors that could improve adaptivity in future collaborative learning systems.
The study was conducted in a public middle school in the southeastern United States in Spring 2024. This IRB-approved study included 128 students across six class periods in a science class. Out of these, 100 consented to participate in the research, and 97 reported their demographic information: 49 identified as girls, 46 as boys, one as non-binary, and one preferred not to disclose. Further, 38 students identified as Asian, 34 as White, 20 as Black/African American, six as Hispanic/Latinx, three as Native American, ifve as self-described, and three preferred not to disclose. The average age was 11.7 years (SD = 0.48).
The primary objective of this study was to engage students in core concepts of artificial intelligence (AI) and computer science (CS) through the hands-on development of conversational agents [39]. The classroom study contained a 10-hour learning module covering AI fundamentals, hands-on activities, and chatbot development. The instructional content was aligned with AI Five Big Ideas in K12 [40]. Specific learning goals included (1) understanding the role of data in AI systems and determining appropriate datasets for specific AI applications; (2) describing how datasets create representations of the world for reasoning tasks; (3) training and evaluating classification and prediction models while examining their accuracy on new inputs; and (4) creating chatbots and developing natural interactions. Each learning module was designed to build students’ AI literacy progressively.
During collaborative chatbot development, students were randomly paired to create chatbots on science topics of their choice (e.g., the water cycle, climate, and living organisms). Each pair collaborated over three class sessions (40 minutes each) to develop their chatbot. We captured audio/video recordings of these collaborative sessions using recording software from their laptop. Humans transcribed the recordings through the Rev service. The resulting dialogue dataset contains 121 collaboration sessions from 47 student pairs (94 individuals). Each session contains an average of 278 utterances (SD = 108.7).
The learning environment, AMBY (Figure 1), is a graphical interface designed for middle school students to create conversational agents and learn about AI [41]. The environment enables students to test example agents, edit agents, or create new agents. For instance, a student group might create an agent on the topic of “climate change,” where they define intents (e.g., “impact on oceans”), input training phrases (e.g., “Does climate change impact the oceans?”), and write corresponding responses (e.g., “Climate change impacts oceans through sea-level rise and ocean acidification.”). Students can test their agents by having conversations and adjusting the chatbot’s voice to personalize the interaction.
Students construct and refine intents (i.e., basic chatbot development units representing the goal or purpose of a user message). In the development panel (Figure 1, left), students can add, edit, or delete training phrases and responses. They submit training requests by clicking “train the AI,” which updates the chatbot’s intent classification model based on the provided training data. Students worked with a partner at the same computer on iterating on the chatbot design, following the pair programming paradigm [42], prompting spontaneous collaboration, dialogue, and joint decision-making.
The system collects 23 types of timestamped user interaction logs (e.g., adding or deleting training phrases, submitting AI training requests, chat messages). Across 121 recorded collaborative sessions, students submitted an average of seven intent training requests per session, and we used the content of training phrases they submitted each time as measures of project quality (Section 3.4).
To model relationships between system logs, student collaborative dialogue, and project outcomes, we prepared a multimodal dataset (Figure 2). Following Borchers et al. [18], we synchronized dialogue transcripts with log data by aligning timestamps between the datasets, yielding an accuracy of 1 second. For each intent submission, we extracted its submitted training phrases for outcome measurement and defined intent-working segment, which is the time window from when students began working on the intent until submission (based on logs). Based on an ad-hoc decision, we manually examined the segments for accuracy and removed segments that were excessively long (> 12.5 minutes, 90 percentile of the dataset) based on considerations of context length during embedding. This preprocessing step is not expected to introduce bias, as our three outcome measures were virtually uncorrelated with dialog length as measured in the total number of words (| | < 0.05 ) and total number of turns (| | < 0.07 ). The average segment duration is 193 seconds. We then extracted the synchronized student dialogues and system logs within these windows for downstream analysis (Section 4).
Our unit of analysis is the intent, a core chatbot element that handles the recognition of user queries. Students averaged seven intent training submissions per 30-minute collaborative session and 17 total submissions. We measured intent quality through three key outcomes chosen based on the learning objectives, learning curve analyses, and correlation with final project scores.
Training Phrase Count measures the number of phrases input by students for training the chatbot to recognize an intent. This metric represents productivity and engagement in the iterative training process: more phrases typically indicate greater efort in improving the chatbot’s performance. The relationship between chat-based activity and engagement has also been observed in past research [19]. Lexical Density captures the proportion of content words (nouns, adjectives, verbs, and adverbs) to total words in the provided phrases, which assesses linguistic richness since efective training requires substantive content over filler words [ 43, 44]. Lexical Variation measures language diversity through the ratio of unique content words to total content words [43], important because varied phrases help chatbots recognize diferent user expressions. Higher variation indicates more robust training input.
These project quality metrics align with key AI learning objectives [40, 39]. For instance, students learn that simply adding similar word permutations (i.e., high training phrase quantity with little information richness, such as ['hat', 'hats', 'HATS']) is inefective for chatbot programming. As students iteratively generate, refine, and diversify training phrases, they engage in active problemsolving and knowledge construction [45, 46]. Lexical diversity is also an important predictor of academic performance in tasks requiring writing and speaking [47, 48].
Validation. We validated our outcomes through learning curve analyses [49] and correlations with three final project grades (i.e., expert-rated score aggregated across 11 dimensions using a rubric, expert-rated training phrase score, and end-user satisfaction evaluated by three independent annotators). Learning curves can demonstrate whether students improve on our outcome measurements as they receive indirect feedback from AMBY on their chatbot testing performance. Figure 3 shows improvements across all metrics over time as students learned to add more phrases and improve lexical choices in the training data for better intent recognition (with the clearest improvement in lexical variation). For correlations, while none of our intent-level measures significantly correlated with overall expert ratings–likely due to the broad scope of the overall assessment, training phrase count significantly correlated with expert-rated training phrase scores (r(45) = 0.5, p < .001), and lexical density correlated with end-user satisfaction score (r(45) = 0.36, p = .013), indicating richer linguistic content improves end-user experience. Our measures showed appropriate separation: moderate correlation between lexical density and variation ( = 0.28 ) but minimal correlation with training phrase count ( = 0.01 and = −0.03 , respectively).
Log Data Features To capture students’ system interactions during chatbot development, we engineered log features reflecting engagement, timing, and strategy as inputs for our machine learning models. We computed timing-related features (six features) capturing temporal dynamics: mean, standard deviation, minimum, maximum, median, and IQR of time between actions. Event-based features (2124 features) characterized interaction patterns through event type counts and bigram/trigram sequences of the 31 unique event types (e.g., [test-chatbot, create-new-intent, add-phrase]). This approach of using log patterns to predict process outcomes aligns with common EDM methods where log-based learning rates relate to instructional events [50, 51].
Language Features in Dialogue Transcripts To capture student dialogue diferences, we computed numerical representations of student dialogue transcripts within a certain segment using state-ofthe-art embedding models. Our early experiments showed Sentence-BERT (SBERT) embeddings [52] from bert-base-uncased performed comparably to OpenAI’s text-embedding-3-large model [53]. Given computational eficiency and open-source considerations, we used SBERT’s 768-dimensional embeddings as input features for predictive modeling, following established EDM methods [20].
We employed a feedforward neural network for regression tasks [20], aiming to predict training phrase quality based on log data features and embeddings. The model architecture consisted of 2-4 hidden layers with ReLU activations and dropout regularization. Hyperparameters were optimized using grid search with five-fold student-level cross-validation over the following parameter space: Hidden layer configurations: [256,128], [512,256,128], [1024, 512, 256,128] and Dropout rates: 0.0, 0.025, 0.05, 0.1, 0.3, 0.5. Model training was conducted using the Adam optimizer, and performance was evaluated using Mean Absolute Error (MAE) and AUC. Early stopping was implemented with a patience threshold of two epochs to prevent overfitting. The code is in the study’s digital appendix. 2
After model training terminated, we evaluated models on a held-out test set (33% of the data), reporting predictive performance across three feature configurations: log only, dialogue only, and combined. Using bootstrapped resampling, we computed 95% AUC confidence intervals (CIs). Based on average cross-validation AUC across folds, the best model was evaluated on the held-out test set.
We evaluated models trained on dialogue data and on log data to predict training phrase quality. Table 1 summarizes the holdout AUC. Training phrase count was better predicted by log-based features (AUC = 0.8053) than dialogue-based features (AUC = 0.5971), suggesting system interaction logs (e.g., frequency and timing of edits) are stronger indicators of phrase production. Conversely, lexical density was better captured by the dialogue-only model (AUC = 0.6551 vs. 0.5112), indicating student talk features capture the richness of their written training phrases. For lexical variation, both models performed modestly (log: AUC = 0.6016, dialogue: AUC = 0.5260), with log features showing a slight advantage.
To answer RQ2, we performed early fusion, concatenating log, and dialogue features before passing them to the model [24]. The results are in Table 1 (“combined” rows). Comparing the best single-modality 2https://github.com/conradborchers/collaboration-edm25 results, for training phrase count, fusion improved AUC from 0.8053 (log-only) to 0.8301.3 However, for lexical density of training phrases, the combined model (AUC = 0.5700) did not outperform the dialogue-only model (AUC = 0.6551), suggesting that log-based features provide limited insight into the content richness of learner input. For lexical variation of training phrases, fusion yielded a slight AUC increase (0.6089 vs. 0.6016 log-only), but confidence intervals suggest the gain is likely not significant. Overall, multimodal fusion yields the most substantial gain for predicting training phrase count, slightly improves the prediction of lexical variation of training phrases, but does not help predict lexical density of training phrases. Hence, the utility of combining modalities is outcome-dependent.
Our findings reveal the diferential efectiveness of dialogue and log data in predicting project quality on the process level. The superior performance of log-derived features in predicting training phrase count aligns with past work on using clickstream data to model learning trajectories [25], indicating behavioral engagement metrics provide critical insights into productivity and task completion. The stronger performance of dialogue features in predicting lexical density indicates the spoken discourse appears to translate into more content-rich chatbot training phrases. This builds on research demonstrating the value of transcribed speech in detecting self-regulated learning strategies [20] and suggests that students’ verbal articulation during collaboration may influence the quality of their computational artifacts [54].
Overall, log-data-based features were more predictive of engagement, while dialogue-based features were more indicative of the linguistic characteristics of chatbot training phrases. Future research should explore how these models learn associations between collaborative actions and outcomes, potentially incorporating interpretability methods to examine feature importance more precisely [55].
Multimodal fusion demonstrated mixed results in enhancing predictive accuracy. The improved prediction of training phrase count through combined features suggests that integrating behavioral and linguistic cues provides a more comprehensive understanding of student engagement. However, for the lexical density of training phrases, the combined model underperformed compared to the dialogue model, suggesting that adding log features may interfere with or dilute the signal from dialogue features for this task. The cross-modal interaction has also been noted in past work [56], emphasizing that fusion requires careful consideration of interference. The modest gains in predicting lexical variation align with past work [25, 31]: the efectiveness of multimodal fusion depends on the prediction task.
The results reinforce the novelty of our study’s focus on open-ended project-based learning in K-12 AI education, highlighting that predicting linguistic quality in this context presents unique challenges and opportunities compared to traditional educational data mining tasks. It is evident from our results 3As one attentive reviewer pointed out, counting ’add-training-phrase’ logs might straightforwardly reveal the number of training phrases. We performed an ablation study, recomputing these accuracies without features dependent on ’add-trainingphrase’, ’delete-phrase’, and ’add-response’. For completeness outcome, AUCs changed to 0.6865 (log only) and 0.6759 in the combined modalities, respectively. that straightforward associations between collaborative dialog and process outcomes during learning are challenging to isolate in our context (as one might conceive that predicted log inputs may be clearly visible and explicitly mentioned in dialog, and hence straightforward to predict; from our experience, these cases rarely occur). This contributes to the ongoing discourse on the role of multimodal data for learning process prediction [26, 19, 20].
The findings highlight the importance of selecting appropriate data modalities for specific outcomes, contributing to multimodal learning analytics [26, 24]. In K-12, AI learning involving iterative chatbot design, log data, and dialogue provides complementary insights into productivity and AI understanding.
For educators assessing collaborative learning, log features ofer insights into engagement and productivity [25], while dialogue features reveal conceptual understanding and self-regulation [19, 20]. However, our mixed multimodal fusion results suggest that careful feature selection and tuning are essential to maximize predictive accuracy. These insights inform future research on collaborative AI literacy tools. With suficient accuracy, our models could monitor engagement and identify when students need scafolding [ 34, 35]. Dialogue analysis could help educators understand students’ AI concept mastery, enabling targeted feedback and aligning with growing interest in natural languageinformed interventions [20, 27, 22]. Future work should evaluate automated transcription feasibility in real classrooms, balancing efectiveness with privacy considerations and teacher preferences [ 57].
First, our data collection was limited to a single middle school science class with a particular collaborative
task, limiting generalizability. Second, our predictive models showed only modest performance in
predicting lexical variation of training phrases, suggesting that the features and modeling techniques
used might not fully capture the complexity of this outcome. These models are not deployment-ready;
their current utility might lie in ofline teacher analytics to inform instruction [ 22]. Future studies
could explore additional features, such as discourse-level linguistic properties or turn-taking dynamics
[
This study examines the accuracy of dialogue and log data in predicting process-level collaboration outcomes in middle school chatbot development. Log features best predict student productivity, while dialogue-based features are better suited to capturing training phrase richness. Multimodal fusion improves predictions for training phrase count and, to a lesser extent, lexical variation, highlighting the outcome-dependent nature of multimodal models. Our research contributes to the field of multimodal learning analytics by demonstrating the potential and limitations of integrating dialogue and log data to predict collaborative learning processes and deploying related models in real-time educational settings. We also contribute insights into the interplay between student interactions and linguistic outputs in natural language programming tasks. By leveraging both dialogue and log data, educators and researchers may gain a more holistic understanding of student learning processes, informing the development of AI-powered collaborative learning tools in K-12 settings.
This research was supported by the National Science Foundation through grant DRL-2048480. We thank Christine Fry Wise for her copy-editing of the manuscript.
During the preparation of this work, the author(s) used ChatGPT, Grammarly in order to: Grammar and spelling check, Paraphrase and reword. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. [15] F. Morshed Fahid, X. Tian, A. Emerson, J. B. Wiggins, D. Bounajim, A. Smith, E. Wiebe, B. Mott, K. Elizabeth Boyer, J. Lester, Progression trajectory-based student modeling for novice block-based programming, in: Proceedings of the 29th ACM Conference on User Modeling, Adaptation and Personalization, 2021, pp. 189–200. [16] S. Marwan, J. Jay Williams, T. Price, An evaluation of the impact of automated programming hints on performance and learning, in: Proceedings of the 2019 ACM Conference on International Computing Education Research, 2019, pp. 61–70. [17] A. Emerson, F. J. Rodríguez, B. Mott, A. Smith, W. Min, K. E. Boyer, C. Smith, E. Wiebe, J. Lester, Predicting early and often: Predictive student modeling for block-based programming environments., Proceedings of International Conference on Educational Data Mining (EDM) (2019). [18] C. Borchers, J. Zhang, R. S. Baker, V. Aleven, Using think-aloud data to understand relations between self-regulation cycle characteristics and student performance in intelligent tutoring systems, in: Proceedings of the 14th Learning Analytics and Knowledge Conference, 2024, pp. 529–539. [19] C. Borchers, K. Yang, J. Lin, N. Rummel, K. R. Koedinger, V. Aleven, Combining dialog acts and skill modeling: What chat interactions enhance learning rates during ai-supported peer tutoring?, in: Proceedings of the International Conference on Educational Data Mining (EDM), 2024. [20] J. Zhang, C. Borchers, V. Aleven, R. S. Baker, Using large language models to detect self-regulated learning in think-aloud protocols, in: Proceedings of the International Conference on Educational Data Mining (EDM), 2024. [21] D. Spikol, E. Rufaldi, M. Cukurova, Using multimodal learning analytics to identify aspects of collaboration in project-based learning, in: CSCL’17: The 12th International Conference on Computer Supported Collaborative Learning, volume 1, International Society of the Learning Sciences., 2017, pp. 263–270. [22] L. Yan, V. Echeverria, Y. Jin, G. Fernandez-Nieto, L. Zhao, X. Li, R. Alfredo, Z. Swiecki, D. Gašević, R. Martinez-Maldonado, Evidence-based multimodal learning analytics for feedback and reflection in collaborative learning, British Journal of Educational Technology 55 (2024) 1900–1925. [23] J. K. Olsen, K. Sharma, N. Rummel, V. Aleven, Temporal analysis of multimodal data to predict collaborative learning outcomes, British Journal of Educational Technology 51 (2020) 1527–1547. [24] K. Sharma, M. Giannakos, Multimodal data capabilities for learning: What can multimodal data tell us about learning?, British Journal of Educational Technology 51 (2020) 1450–1484. [25] S. Karumbaiah, C. Borchers, T. Shou, A.-C. Falhs, P. Liu, T. Nagashima, N. Rummel, V. Aleven, A spatiotemporal analysis of teacher practices in supporting student learning and engagement in an ai-enabled classroom, in: International Conference on Artificial Intelligence in Education, Springer, 2023, pp. 450–462. [26] W. Chango, J. A. Lara, R. Cerezo, C. Romero, A review on data fusion in multimodal learning analytics and educational data mining, Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 12 (2022) e1458. [27] A. Scarlatos, A. Lan, Exploring knowledge tracing in tutor-student dialogues, arXiv preprint arXiv:2409.16490 (2024). [28] E. Walker, N. Rummel, K. R. Koedinger, Adaptive intelligent support to improve peer tutoring in algebra, International Journal of Artificial Intelligence in Education 24 (2014) 33–61. [29] J. Costley, How system functionality improves the efectiveness of collaborative learning, Interactive Learning Environments 30 (2022) 971–983. [30] J. Roschelle, S. D. Teasley, The construction of shared knowledge in collaborative problem solving, in: Computer supported collaborative learning, Springer, 1995, pp. 69–97. [31] V. Echeverria, L. Yan, L. Zhao, S. Abel, R. Alfredo, S. Dix, H. Jaggard, R. Wotherspoon, A. Osborne, S. Buckingham Shum, et al., Teamslides: a multimodal teamwork analytics dashboard for teacherguided reflection in a physical learning space, in: Proceedings of the 14th Learning Analytics and Knowledge Conference, 2024, pp. 112–122. [32] Q. Zhou, W. Suraworachet, M. Cukurova, Detecting non-verbal speech and gaze behaviours with multimodal data and computer vision to interpret efective collaborative learning interactions, Education and Information Technologies 29 (2024) 1071–1098. [33] M. Abdelshiheed, J. K. Jacobs, S. K. D’Mello, Aligning tutor discourse supporting rigorous thinking with tutee content mastery for predicting math achievement, in: Proceedings of the 25th International Conference on Artificial Intelligence in Education (AIED’24), Recife, Brazil, 2024. [34] A. M. Latham, K. A. Crockett, D. A. McLean, B. Edmonds, K. O’shea, Oscar: An intelligent conversational agent tutor to estimate learning styles, in: International conference on fuzzy systems, IEEE, 2010, pp. 1–8. [35] N. Matsuda, E. Yarzebinski, V. Keiser, R. Raizada, G. J. Stylianides, W. W. Cohen, K. R. Koedinger, Learning by teaching simstudent–an initial classroom baseline study comparing with cognitive tutor, in: International Conference on Artificial Intelligence in Education (AIED 2011), Springer, 2011, pp. 213–221. [36] X. Tian, N. Lubold, L. Friedman, E. Walker, Understanding rapport over multiple sessions with a social, teachable robot, in: International Conference on Artificial Intelligence in Education (AIED 2020), Springer, 2020, pp. 318–323. [37] A. Weerasinghe, A. Mitrovic, M. Van Zijl, B. Martin, Evaluating the efectiveness of adaptive tutorial dialogues in database design, in: Proceedings of the 18th International Conference on Computers in Education, 2010, pp. 33–40. [38] S. D’Mello, A. Olney, C. Williams, P. Hays, Gaze tutor: A gaze-reactive intelligent tutoring system,
International Journal of Human-Computer Studies 70 (2012) 377–398. [39] Y. Song, G. A. Katuka, J. Barrett, X. Tian, A. Kumar, T. McKlin, M. Celepkolu, K. E. Boyer, M. Israel, Ai made by youth: A conversational ai curriculum for middle school summer camps, in: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Innovative Applications of Artificial Intelligence Conference and Thirteenth AAAI Symposium on Educational Advances in Artificial Intelligence, 2023. [40] D. Touretzky, C. Gardner-McCune, F. Martin, D. Seehorn, Envisioning ai for k-12: What should every child know about ai?, in: Proceedings of the AAAI conference on artificial intelligence, volume 33, 2019, pp. 9795–9799. [41] X. Tian, A. Kumar, C. E. Solomon, K. D. Calder, G. A. Katuka, Y. Song, M. Celepkolu, L. Pezzullo, J. Barrett, K. E. Boyer, et al., Amby: A development environment for youth to create conversational agents, International Journal of Child-Computer Interaction 38 (2023) 100618. [42] B. Hanks, S. Fitzgerald, R. McCauley, L. Murphy, C. Zander, Pair programming in education: A literature review, Computer Science Education 21 (2011) 135–173. [43] J.-y. Kim, Predicting l2 writing proficiency using linguistic complexity measures: A corpus-based study., English teaching 69 (2014). [44] X. Lu, The relationship of lexical richness to the quality of esl learners’ oral narratives, The
Modern Language Journal 96 (2012) 190–208. [45] W. Park, H. Kwon, Implementing artificial intelligence education for middle school technology education in republic of korea, International Journal of Technology and Design Education 34 (2023) 109–135. doi:10.1007/s10798- 023- 09812- 2. [46] K. Lakkaraju, T. Hassan, V. Khandelwal, P. Singh, C. Bradley, R. Shah, F. Agostinelli, B. Srivastava, D. Wu, Allure: a multi-modal guided environment for helping children learn to solve a rubik’s cube with automatic solving and interactive explanations, Proceedings of the AAAI Conference on Artificial Intelligence 36 (2022) 13185–13187. doi: 10.1609/aaai.v36i11.21722. [47] G. Yu, Lexical diversity in writing and speaking task performances, Applied linguistics 31 (2010) 236–259. [48] B. Kondal, Efects of lexical density and lexical variety in language performance and proficiency,
International Journal of IT, Engineering and Applied Sciences Research (IJIEASR) 4 (2015) 25–29. [49] K. Rivers, E. Harpstead, K. R. Koedinger, Learning curve analysis for programming: Which concepts do students struggle with?, in: Proceedings of 2016 ACM Conference on International Computing Education Research, volume 16, ACM, 2016, pp. 143–151. [50] M. Chi, K. Koedinger, G. Gordon, P. Jordan, Instructional factors analysis: A cognitive model for multiple instructional interventions, in: Proceedings of the 4th International Conference on Educational Data Mining (EDM), 2011. [51] J. Lin, S. Singh, L. Sha, W. Tan, D. Lang, D. Gašević, G. Chen, Is it a good move? mining efective tutoring strategies from human–human tutorial dialogues, Future Generation Computer Systems 127 (2022) 194–207. [52] N. Reimers, Sentence-bert: Sentence embeddings using siamese bert-networks, arXiv preprint arXiv:1908.10084 (2019). [53] L. Wang, N. Yang, X. Huang, L. Yang, R. Majumder, F. Wei, Improving text embeddings with large language models, arXiv preprint arXiv:2401.00368 (2023). [54] S. Grover, R. Pea, Computational thinking in k–12: A review of the state of the field, Educational researcher 42 (2013) 38–43. [55] A. Condor, Z. Pardos, Explainable automatic grading with neural additive models, in: International
Conference on Artificial Intelligence in Education, Springer, 2024, pp. 18–31. [56] J. Hessel, L. Lee, Does my multimodal model learn cross-modal interactions? it’s harder to tell than you might think!, Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) (2020). doi:10.18653/v1/2020.emnlp- main.62. [57] K. B. Yang, C. Borchers, A.-C. Falhs, V. Echeverria, S. Karumbaiah, N. Rummel, V. Aleven, Leveraging multimodal classroom data for teacher reflection: Teachers’ preferences, practices, and privacy considerations, in: European Conference on Technology Enhanced Learning, Springer, 2024, pp. 498–511. [58] Y. Ma, Multimodal Modeling of Collaborative Learning with Adaptive Data Fusion, Ph.D. thesis, 2023. [59] C. Cohn, E. Davalos, C. Vatral, J. H. Fonteles, H. D. Wang, M. Ma, G. Biswas, Multimodal methods for analyzing learning and training environments: A systematic literature review, arXiv preprint arXiv:2408.14491 (2024). [60] M. Wixon, R. S. d. Baker, J. D. Gobert, J. Ocumpaugh, M. Bachmann, Wtf? detecting students who are conducting inquiry without thinking fastidiously, in: User Modeling, Adaptation, and Personalization: 20th International Conference, UMAP 2012, Montreal, Canada, July 16-20, 2012. Proceedings 20, Springer, 2012, pp. 286–296.