Our Adversarial Training Paper has been Accepted by AISTATS 2021
Our Adversarial Training Paper has been Accepted by AISTATS 2021. We propose a learning-to-learn method to improve adversarial training on a variety of tasks.
Our BERT Calibration Paper has been Accepted by EMNLP 2020
We propose to improve BERT model calibration via on-manifold smoothing and off-manifold smoothing.
Two Papers are Accepted by ICML 2020
Our recent works on transformer-based point process model and smooth policy of reinforcement learning are accepted by ICML 2020!
BOND is Accepted by KDD 2020
Our recent work on distantly supervised NER is accepted by KDD2020. We are closing the GAP between full supervision and distant supervision!
Two Papers are Accepted by ACL 2020
Our recent works on BERT fine-tuning and multidomain NMT are accepted by ACL2020!
State-of-the-art Results on GLUE Benchmark
SMART model achieves 89.9 on GLUE, even better than T5!