A hybrid recommendation system for Chinese character components using a dual-tower neural network and a deterministic algorithm

The accelerating globalization necessitates innovative solutions for Chinese character acquisition, particularly for non-native learners at the foundational stage. These learners often struggle with the structural logic of component-based logograms, a critical bottleneck that rote memorization fails to address.Existing mobile learning tools predominantly emphasize rote memorization, failing to establish systematic understanding of character formation principles. This study presents Han Zi Builder—a Unity3D-based mobile game integrating component assembly mechanics with a hybrid AI system—to bridge this critical pedagogical gap. The framework synergizes deterministic rule validation (98.2% radical compatibility accuracy) and a dual-tower neural network (768d BERT embeddings with dynamic negative sampling), supported by a meticulously curated 141-radical/803-character database encoding 12,345 valid combinations.Controlled experiments (N = 127 learners) demonstrate three key advancements: (1) 80.4% recommendation F1 score (95%CI:78.9–81.7%), outperforming rule-based baselines by 29.5% (p < 0.001); (2) 30.7% sustained engagement increase measured through playtrace analysis (χ²=25.33); (3) 78ms mobile inference latency achieved via ONNX quantization (model size 202KB). The system’s cognitive loop design—featuring real-time visual feedback (zooming/highlighting)—significantly enhances learners’ morphological awareness. This is evidenced by a 36.3% improvement in character decomposition accuracy (Cohen’s d = 1.47), a key cognitive skill that precedes long-term memory retention. Furthermore, this enhanced awareness strongly correlates with HSK progression (r = 0.71, p < 0.01), suggesting that the proposed framework provides a solid foundation for sustained language acquisition.Future research will implement transformer-based behavior modeling to advance adaptive personalization. While this work focuses on optimizing for mobile accessibility using a dual-tower architecture, future research will explore lightweight Transformer-based models to capture dynamic contextual semantics for more complex character structures.