tree-mlp

This paper proposes CNM, a lightweight augmentation that injects discrete compositional structure of Chinese characters into BERT via Ideographic Description Sequences, improving performance on rare and out-of-vocabulary characters while preserving general NLU accuracy.