SGNPP: A Graph Neural Network-Based Framework for Japanese Grammar Modeling and Adaptive Learning Path Optimization

Abstract

Japanese grammar, with its hierarchical structures, flexible word order, and context-driven usage, presents unique challenges for both computational analysis and personalized language learning. Existing methods, including rule-based parsing and sequence-oriented models, often overlook the interconnectedness among grammatical constructs, leading to fragmented analysis and static, one-size-fits-all learning paths. Traditional approaches frequently struggle to capture long-range dependencies and prerequisite relationships between grammar patterns, limiting their ability to guide effective learning progressions. These shortcomings hinder adaptive curriculum design and individualized progression tracking for learners. To address these issues, this study introduces the Syntax Graph Neural Path Planner (SGNPP), which models Japanese grammar as a structured graph where nodes represent morphemes, bunsetsu units, and grammar patterns. At the same time, the edges encode syntactic and prerequisite relationships. Using graph neural networks, SGNPP performs grammar recognition, error diagnosis, and learning dependency modeling. Reinforcement signals are integrated to optimize personalized learning paths, balancing mastery gain with efficiency and cognitive load. The proposed method enables adaptive sequencing of grammar lessons, real-time feedback, and dynamic instruction adjustment based on learner performance. The experiments show that SGNPP has a grammatical identification accuracy of 92%, outperforming CILS (78%), FIAS (82%), and TTS (86%). A practical learning trajectory study shows that SGNPP has 86% mastery of advanced grammar. Quantitative results show that SGNPP is an effective and customized strategy for learning Japanese and improves computational grammar instruction.