Skip to content

5 projects

Work

Two research projects, two tools, one hackathon.

Lead project

The Tree-MLP encodes a character's component tree and its output is fused into the standard BERT embedding before the unchanged transformer backbone.

Read this one first

CNM-BERT

ResearchUnder review2026

BERT treats 樹 as an arbitrary ID and throws away the components a reader sees instantly. CNM-BERT feeds that structure back in and lifts out-of-vocabulary accuracy by 9.8 F1 for about 5% more training time.

PyTorchBERTTree-MLPIDSCLUE
Read the write-up