Prefix Generator for Low-resource Event Extraction, PlusLab @ UCLA

• Incorporated useful external information, such as syntax trees, into generative models, which have been important for event extraction tasks because of their flexibility and efficiency.
• Proposed Prefix Generator, which encodes external information (e.g. Abstract Meaning Representation Graph, Optimus robust representation) by pre-training and mapping representations into prefixes in encoder-decoder models.
• Our method proved to be generally applicable and especially effective in low-resource settings.
• This work will submit to ACL 2023.

Domain Relabeling for Subpopulation Shift, IRIS Lab @ Stanford

• Analyzed spurious correlations caused by poor quality domain labels in current approaches to address subpopulation shifts between training and testing distributions.
• Proposed integrating the candidate features in datasets’ metadata table to get higher-quality domain labels.
• Built reinforcement learning framework that utilizes downstream task performance feedback of each metadata feature to optimize the domain labels iteratively.
• Our method leads to improved worst-group performance in real-world datasets covering fields such as healthcare and weather forecasting.

Reviewing Test Protocols of Distantly Supervised Relation Extraction, THUNLP @ Tsinghua

• Examined two popular relation extraction datasets (NYT10 and Wiki20) for annotation errors due to distant supervision methods adopted.
• Proposed an improved relation ontology and adopted data-cleaning, constructed manually-annotated test sets for NYT10 and Wiki20, correcting 53% wrong labels in NYT10.
• Analyzed performance differences of competitive models on manually-annotated and distantly supervised datasets.
• Our conclusion sheds light on the importance of a more accurate evaluation for relation extraction research.
• Paper Manual Evaluation Matters: Reviewing Test Protocols of Distantly Supervised Relation Extraction published in ACL Findings 2021.