01 · NATURAL LANGUAGE PROCESSING
CNM-BERT: A Drop-In Structural Embedding for Chinese Characters via Ideographic Description Sequences
Token-based encoders treat a Chinese character as an arbitrary identifier and discard the recursive structure a reader sees immediately. CNM parses Ideographic Description Sequences into trees, encodes them with a recursive Tree-MLP, and fuses the result into BERT without touching the backbone. On out-of-vocabulary characters it gains 9.8 Structure F1 over the strongest baseline, for about 5% more training time.
- Authors
- Liqian Yan · Thomas Sing-wing Wu (equal contribution)
- Venue
- ACL Rolling Review 2027
- Status
- Submitted, decision pending
02 · COMPUTATIONAL ECONOMICS
The AIFE Engine: A Computational Framework to Quantify and Forecast the Great Labor Reallocation
One hundred million job postings from 2015 to 2024, scored for AI intensity by a distilled language model that was calibrated against 9,500 human labels. The resulting firm-by-year panel is used to estimate what adoption does to the composition of hiring, and then to forecast where it goes next with distribution-free prediction intervals.
- Authors
- Liqian Yan · Yintong Chen · Zichun Qiu
- Venue
- Regeneron ISEF 2026
- Status
- Finalist · work ongoing