Limitations of Early Teaching Agents
Pedagogical Agents (PA), also known as “teaching agents,” are a key component of Intelligent Tutoring Systems (ITS). Teaching agents represent an important research direction for the application of artificial intelligence technology in education. They are designed as anthropomorphic characters (simulating teachers, learning partners, etc.) to enhance the social attributes of ITS, thereby increasing learning interactions and engagement, improving student motivation, and providing personalized teaching and tutoring. The ITS behind the PA is essentially a rule-based expert system that relies on a predefined subject knowledge base and teaching strategies to provide guidance and feedback to students, offering high accuracy. However, teaching a computer to possess teaching intelligence through manually programming subject knowledge and teaching strategies is very complex and time-consuming. Therefore, ITS struggles to update quickly to reflect the latest knowledge and teaching strategies, which limits the scalability and adaptability of PA, affecting its widespread application.
Exploration of Generative Artificial Intelligence in Teaching
Currently, ChatGPT is being used as a general-purpose teaching agent, serving as a private tutor for learners and a teaching assistant for teachers. However, in actual teaching applications, generative artificial intelligence like ChatGPT has not performed well. The prominent drawbacks include sometimes providing incorrect explanations when answering student questions, which can mislead students; lengthy analyses that increase cognitive load for students; and directly providing answers that encourage cheating.
Researchers have found that the poor performance of generative artificial intelligence directly used in teaching practice is due to the fact that the underlying large models are not specifically developed for educational purposes. Therefore, it is necessary to adjust and optimize the large models for teaching, allowing generative artificial intelligence to interact with students like excellent human teachers. Prompt engineering is the simplest and most commonly used method to optimize large models. It requires teachers to write a set of instructions about good teaching behaviors in natural language as prompts, which the generative artificial intelligence uses to produce corresponding teaching behaviors and content. However, this method has limited effectiveness and does not fundamentally correct the aforementioned issues of the large model. Fine-tuning is the second method for optimizing large models. It involves additional training of the large model, allowing it to learn teaching-related knowledge from supplementary teaching data, understand effective teaching behaviors and strategies, and acquire teaching-related knowledge and skills from the ground up, enabling it to interact with students in a manner more aligned with teaching principles. However, the lack of high-quality teaching datasets and effective evaluation standards makes fine-tuning educational large models far from perfect. Nevertheless, exploration has not ceased; some teachers have found a third method to enable generative artificial intelligence to serve the classroom.
Implementation of Generative Pedagogical Agents
Currently, most generative artificial intelligence platforms support users in creating specific-function agents (AI Agents) with zero-code or low-code methods. Teachers have discovered that agents created in this way can also support their teaching. This represents a new type of teaching agent, and to differentiate it from traditional teaching agents, it is named: Generative Pedagogical Agent (GPA).
GPA uses preset instructions and teaching materials to guide large generative artificial intelligence models like GPT-4 to support student learning in specific subjects. This employs Retrieval-Augmented Generation (RAG), which is the third method to optimize the application of large models in education. RAG enhances generative artificial intelligence by combining prompts with knowledge base content to exhibit better teaching behaviors and content output. The prompts here are preset instructions, which serve as clear indications of the agent’s teaching behaviors. The knowledge base consists of teaching materials provided by teachers, clarifying the content of teaching for the agent. The combination of the two, leveraging the large model’s powerful understanding and generalization capabilities, can enhance the teaching ability of the agent. The specific implementation process is that when a student poses a question, the agent combines the preset instructions with the student’s question into a new instruction, retrieves relevant content from the knowledge base using the new instruction, and then constructs a context-rich and structurally sound prompt from the new instruction and the retrieved content to submit to the large model for execution. Thus, GPA can better accomplish its teaching tasks. It can be seen that GPA integrates the advantages of traditional teaching agents and generative artificial intelligence, ensuring the accuracy and adaptability of teaching through presetting and generation. However, achieving this requires certain methods and techniques.
These strategies are further establishing “guardrails” for GPA to play a stable role in teaching specific subjects. It is noteworthy that some specialized GPA creation platforms (such as SchoolAI abroad and Xiaoyang Intelligent Agent domestically) have set better “guardrails” to support safe teaching, such as preventing interference from unrelated agents and allowing teachers to view dialogue data between students and agents.
Roles of Generative Pedagogical Agents
Recent studies have found that assigning roles to the large model in prompts can enhance its reasoning ability and content output quality. When creating GPA, the following methods can be used to assign roles to improve its performance.
(1) Subject Role.This is a role assigned to create a GPA for a specific teaching subject, such as a writing mentor for expository texts, a vocabulary coach, or a guide for balancing chemical equations; it can also be created for a conceptual subject, a reading material, or a historical event. In practice, the most common case is for teachers to create GPA mentors based on the content of a unit or a lesson. Teachers only need to specify teaching requirements (teaching steps, methods, etc.) in the preset prompts and upload corresponding teaching materials to easily create this type of GPA. This type of GPA can typically support students’ autonomous learning through full-process teaching, not only answering student questions but also helping students learn by asking them questions about the subject.
(2) Functional Role.If the GPA is not used for full-process teaching but rather as a teaching assistant for a specific teaching segment, a GPA can be created to achieve specific teaching functions, such as a chemistry bond Q&A assistant (specifically answering various questions about chemical bonds), a trigonometric function exercise generator (generating problems for students to practice based on their requests and providing feedback), or a current exercise review assistant (individually reviewing students’ assignments on the “current” lesson). The purpose of this type of GPA is to assist teacher-led instruction. Thus, creating this type of GPA is simpler than for subject roles. For instance, in creating an exercise review assistant, teachers only need to specify review rules in the preset prompts and upload the exercises and answers that need to be reviewed.
(3) Character Role.This involves creating GPA roles for characters related to the teaching content. We can have GPA play real characters, such as a scientist, writer, or historical figure, or have GPA play characters created in literary works, such as Shylock from Shakespeare’s Merchant of Venice, or even have GPA play anthropomorphized roles like atoms or cells. This type of GPA supports students in immersive extended learning by creating scenarios, mainly providing rich materials and appropriate explanations. Creating this type of GPA may be simpler. First, there is no need to specify complex teaching rules in the preset prompts; it is sufficient to provide requirements for the correctness, difficulty, and length of the output content to suit the students’ levels. Second, the underlying generative artificial intelligence model of GPA can already answer most questions, so the demands for knowledge base materials are not high; only content that the large model is unfamiliar with needs to be prepared, and it is recommended to provide relevant textbook content to help GPA understand the knowledge levels students need to master.
Applications of Generative Pedagogical Agents
(1) Flipped Classroom.The flipped classroom involves students learning new knowledge at home through course materials (videos or other digital resources) and then applying what they have learned in class through collaborative activities, discussions, or problem-solving exercises. In this process, even well-prepared course materials by teachers may still fall into the trap of a “one-size-fits-all” approach. GPAs created based on thematic content can serve as AI mentors to support students in learning more effectively at home, ensuring they come to class better prepared and more engaged in practical activities or discussions. Additionally, in classes with a large number of students, teachers can create GPAs as Q&A assistants to answer personalized questions from students, and distribute exercise generator GPAs to each student for adaptive practice and feedback.
(2) Peer Teaching.Peer teaching utilizes discussions and exchanges between students to promote their understanding and application of concepts, enhancing their classroom participation and learning outcomes. In this process, many factors may hinder effective discussion and communication, such as verbal expression skills, personality, and cognitive gaps. Teachers can create GPAs to serve as “peers” for students who have poor pairing effects, allowing AI peers to engage in adaptive interactions with them. Furthermore, having students play the role of teachers to teach others is a common activity in peer teaching; however, if the “learner” cannot identify the “teacher”‘s mistakes, this may lead to the risk of inaccurate knowledge dissemination. Teachers can create GPAs to play the role of “learner” to facilitate the effective implementation of this activity and avoid similar situations.
(3) POGIL Model.POGIL stands for Process Oriented Guided Inquiry Learning, which is a form of guided inquiry learning where teachers provide guiding learning materials to support students in collaboratively and progressively exploring problems. In this process, students need to take on different cooperative learning roles, while teachers walk around the classroom to provide guidance and support as needed. If there are defects in student group collaboration, teachers can create GPAs to play corresponding roles in the group to assist in collaborative learning; if teachers are overwhelmed, they can create GPAs to act as team coaches to guide the group.
(4) 5E Model.The 5E model is similar to POGIL; in essence, both are scaffolded teaching based on constructivist learning theory, supporting students’ inquiry learning through five activities: Engage, Explore, Explain, Elaborate, and Evaluate. Teachers can create GPAs for each group to serve as process coaches, enhancing guidance for each group’s activities, including providing task questions at the start of each activity, assisting with common questions during the process, and adaptively guiding the subsequent activities by receiving group activity outcomes and providing evaluations and feedback.
Future Prospects of Teaching Agents
Early PAs appeared as 2D or 3D anthropomorphic figures, but ultimately did not achieve widespread application due to the complexity of technical implementation and limited interactive effects. Now, with generative artificial intelligence, it is possible to create digital avatars with real human appearances—digital humans—while GPT-4 can observe the real world through cameras and interact with humans. If these advanced technologies are combined with GPA, it may represent the future form of GPA—digital human teaching agents. It is foreseeable that when simulated teachers in real human form can communicate with students based on visual input, education will truly undergo revolutionary changes.
Authors │ Zhang Yijiang, Dai Haijun, Luo Tailiang, Lan Yong
Affiliation │ Chongqing Ju Kui Middle School
Source │“Information Technology Education in Primary and Secondary Schools”, Issue 10, 2024
Click the above image to enter the online subscription page of the magazine
Recently, some teachers received the following manuscript notification and certificate, both of which are counterfeit documents and seals.
Special Reminder! Teachers must not believe it! Do not transfer payment!
↓↓↓
Submissions should be sent to official submission email ([email protected]) and WeChat mini-program “Information Technology Education Magazine for Primary and Secondary Schools“.
Typesetting | Yi Dan Initial Review | Zhang Ying
Re-review | Lu Qiuhong Final Review | Su Jinshu
Like this article? Don’t forget to click “See” to support~
What You Want to SeeWonderful
New Issue Dispatch│Table of Contents for Issue 10, 2024
Welfare is Here! Welcome to Subscribe to the “Information Technology Education” Magazine for 2024
Call for Papers! Welcome to Participate in the 2024 “Digital Empowerment in Primary and Secondary School Teaching Practice” Paper Collection Activity
2024 Collection of Innovative Applications of Information Technology by Primary and Secondary School Students is Now Open
