CNM-BERT: A Drop-In Structural Embedding for Chinese Characters via Ideographic Description Sequences

Chinese characters are traditionally tokenized as atomic IDs in models like BERT, which ignores their internal structure and hurts performance on rare or unseen characters. The Compositional Network Model (CNM) parses Ideographic Description Sequences (IDS) to represent characters as discrete compositional trees, then augments standard Transformer encoders as a drop-in upgrade. This approach adds no heavy pre-training and can be applied to existing BERT models. The authors demonstrate that CNM-BERT improves character-level and downstream task performance, especially for low-frequency and OOV characters. The work suggests that explicit orthographic structure can complement context-based learning in Chinese NLP.