Unsupervised Domain Adaptation for Event Detection via Meta Self-Paced Learning

Abstract

As important events in textual data are usually highly specific in terms of tasks and domains, a change in data distribution would have a significant impact on detection performance. Recent methods addressing unsupervised domain adaptation for event detection task typically extracted domain-invariant representations through combining and balancing various objectives to align the feature space between source and target domains. While effective, these methods are impractical as large-scale language models are drastically growing bigger to achieve optimal performance. To this end, we propose Meta Self-Paced Domain Adaption framework (MSP-DA) that effectively and efficiently alleviates the need for domain-specific hyperparameter tuning. By imitating the train-test dataset split based on the difficulties of source domain's samples, the model is trained through a meta-learning process that learns to weigh the importance of each labeled instance and to balance every alignment objective, simultaneously. Extensive experiments demonstrate our framework substantially improves performance on target domains, surpassing state-of-the-art approaches. Furthermore, we present detailed analyses to validate our method and provide insight into how each domain affects the learned hyperparameters.

Tasks

Reproductions