Abstract
With the increasing attention to the cultivation of legal talents, a new teaching model has been explored through artificial intelligence (AI) technology under educational psychology, which focuses on improving learning initiative, teaching methods, and teaching quality of students. First, the application of AI and deep neural network (DNN) algorithms are reviewed in education, and the advantages and disadvantages of traditional learning material recommendation algorithms are summarized. Then, a personalized learning material recommendation algorithm is put forward based on DNN, together with an adaptive learning system based on DNN. Finally, the traditional user-based collaborative filtering (UserCF) model and lifelong topic modeling (LTM) algorithm are introduced as the control group to verify the performance of the proposed recommendation system. The results show that the best learning rate of model training is 0.0001, the best dropout value is 0.5, and the best batch size is 32. The proposed personalized learning resource recommendation method based on deep learning (DL) still has good stability under various training data scales. The personalized test questions of recommended students are moderately difficult. It is easier to recommend materials according to the acquisition of knowledge points and the practicability of the recommended test questions of students. Personalized learning material recommendation algorithm based on AI can timely feedback needs of students, thereby improving the effect of classroom teaching. Using the combination of AI and DL algorithms in teaching design, students can complete targeted personalized learning assignments, which is of great significance to cultivate high-level legal professionals.
Introduction
Legal education is traditionally rigorous, logical, and often follows a "one-size-fits-all" approach. In many academic contexts, the "teacher-centered" model fails to account for the varying abilities and learning speeds of individual law majors. Personalization at scale remains a significant challenge; while a human tutor can identify specific weaknesses in complex legal subjects like contract law or criminal procedure, doing so for hundreds of students simultaneously is impractical for faculty. Previous attempts using basic recommendation algorithms, such as simple collaborative filtering, often suffer from "cold start" problems or lose accuracy as datasets grow. This research bridges educational psychology and AI to create a system that adapts to a student's unique learning path.
Methodology
The researchers developed an Adaptive Learning System built on a Deep Neural Network (DNN) architecture. The core model uses multiple hidden layers to excavate deep relationships between student behavior and learning materials, mapping interactive records—such as test scores and time spent on topics—to a correlation matrix of knowledge points.
Recognizing that learning is a sequential process—where current mastery depends on previous foundations—the authors integrated a Bi-directional Long Short-Term Memory (Bi-LSTM) network to tackle sequence problems in learning paths. The system architecture is modular, divided into four subsystems: online learning, personalized examination, learning situation analysis (to diagnose student cognition), and resource management.
The model utilizes backpropagation and gradient descent for parameter tuning, specifically employing a Sigmoid activation function to handle output results. The study utilized a real-world dataset consisting of 18,000 records from over 300 law majors across several universities in China.
Key Results and Insights
The authors conducted extensive parameter tuning to find the optimal configuration for the model:
- Optimal Hyperparameters: The best performance was achieved with a learning rate of 0.0001, a dropout value of 0.5 (to prevent overfitting), and a batch size of 32.
- Stability Over Scale: One of the most significant findings was the model's stability. While traditional User-based Collaborative Filtering (UserCF) saw a decline in accuracy as the training set ratio increased, the DNN-based model remained stable and even improved, suggesting deep learning is better suited for complex educational data.
- The "Moderate Difficulty" Sweet Spot: By using Item Response Theory, the system ensures recommended test questions are "moderately difficult." This aligns with the educational psychology concept of the "Zone of Proximal Development," keeping students challenged but not discouraged.
Conclusion and Future Work
While the application of DNNs in legal education is still in its early stages, this study demonstrates that AI-driven personalized recommendation systems can effectively address individual student needs and improve classroom teaching quality. The authors acknowledge that current research is relatively "shallow" and plan to expand the research scope to improve accuracy and further integrate human-computer interaction (HCI) to make the AI feedback loop even more intuitive for students.
Summary updated