Wangda Zhu’s research while affiliated with University of Florida and other places

The ability of large language models (LLMs) to generate code has raised concerns in computer science education, as students may use tools like ChatGPT for programming assignments. While much research has focused on higher education, especially for languages like Java and Python, little attention has been given to K-12 settings, particularly for pseudocode. This study seeks to bridge this gap by developing explainable machine learning models for detecting pseudocode plagiarism in online programming education. A comprehensive pseudocode dataset was constructed, comprising 7,838 pseudocode submissions from 2,578 high school students enrolled in an online programming foundations course from 2020 to 2023, along with 6,300 pseudocode samples generated by three versions of ChatGPT. An ensemble model (EM) was then proposed to detect AI-generated pseudocode and was compared with six other baseline models. SHapley Additive exPlanations were used to explain how these models differentiate AI-generated pseudocode from student submissions. The results show that students’ submissions have higher similarity with GPT-3 than with the other two GPT models. The proposed model can achieve a high accuracy score of 98.97%. The differences between AI-generated pseudocode and student submissions lies in several aspects: AI-generated pseudocode often begins with more complex verbs and features shorter sentence lengths. It frequently includes clear numerical or word-based indicators of sequence and tends to incorporate more comments throughout the code. This research provides practical insights for online programming and contributes to developing educational technologies and methods that strengthen academic integrity in such courses.

Knowledge tracing (KT) is a method to evaluate a student's knowledge state based on their historical problem-solving records by predicting the next answer's binary correctness. Although widely applied to closed-ended questions, it lacks a detailed option tracking method for assessing multiple-choice questions (MCQs). This paper introduces a general option tracking (OT) method that can be seamlessly integrated into deep knowledge tracing (DKT) methods through data processing techniques and network output modules. Using a million-level assignment records of MCQs from a K-12 math learning platform, which includes two types of knowledge components (KC): skill and misconception, we converted five different DKT models into deep option-tracing (DOT) models. Performance metrics demonstrate that OT enhances KT performance and effectively identifies students' future option-selection tendencies. Furthermore, using the best OT model, we extracted students' problem-solving sequence features and learning gains to analyze learning patterns. The results reveal that for beginners in middle school mathematics, consecutive errors in the same skill might lead to greater learning gains. Finally, we applied network analysis to reveal connections between skills based on students' error tendencies. Our work contributes to KT methods and related empirical findings in learning analytics (LA) for knowledge assessment.

Mathematical stories can enhance students' motivation and interest in learning mathematics, thereby positively impacting their academic performance. However, due to resource constraints faced by the creators, generative artificial intelligence (GAI) is employed to create mathematical stories accompanied by images. This study introduces a method for automatically assessing the quality of these multimodal stories by evaluating text‐image coherence and textual readability. Using GAI‐generated stories for grades 3 to 5 from the US math story learning platform Read Solve Create (RSC), we extracted features related to multimodal semantics and text readability. We then analysed the correlation between these features and student engagement levels, measured by average reading time per story (behavioural engagement) and average drawing tool usage per story (cognitive engagement), derived from browsing logs and interaction metrics on the platform. Our findings reveal that textual features such as conjunctive adverbs, sentence connectors, causal connectives and simplified vocabulary positively correlate with behavioural engagement. Additionally, higher semantic similarity between text and images, as well as the number of operators in the stories, is associated with increased cognitive engagement. This study advances the application of GAI in mathematics education and offers novel insights for instructional material design. Practitioner notes What is already known about this topic Mathematical stories can enhance students' motivation and interest in mathematics, leading to improved academic performance. Generative artificial intelligence (GAI) has been increasingly employed to create multimodal educational content, including mathematical stories with accompanying images, to address content creators' resource constraints. Prior readability research has primarily focused on the analysis of text‐based educational content, with less emphasis on the integration and analysis of visual elements. What this paper adds Introduces a novel automated multimodal readability assessment method that evaluates the coherence between text and images and the readability of text in GAI‐generated mathematical stories. Identifies specific story features, such as the more frequent use of three types of conjunctions (adversative conjunctions, common sentence conjunctions and logical conjunctions) and vocabulary simplicity that correlate with student engagement. Implications for practice and/or policy Educators and curriculum developers are encouraged to utilise automated multimodal readability assessment tools to analyse and refine GAI‐generated educational content, aiming to enhance student engagement and learning experience. Suggestions for the design of educational content includes the consideration of identified readability features that correlate with higher engagement. Caution should be exercised in handling the association between images and text considering the cognitive load of the instructional materials.

This study examines public sentiment towards AI in education, focusing on the impact of ChatGPT's launch by OpenAI on November 30, 2022. Analyzing around 80,000 Twitter posts from before and after the launch, we conducted a comprehensive sentiment analysis using a fine-tuned BERT, outperforming traditional methods such as VADER and SVM. We applied an RDD to assess the causal impacts of ChatGPT's introduction on public sentiment track sentiment shifts, highlighting how the introduction of AI technologies like ChatGPT has influenced educational discourse. Our findings reveal significant public sentiment changes post-launch, contributing new insights into AI's role in education and public discourse.