Learning Domain Invariant Word Representations for Parsing Domain Adaptation

摘要

We show that strong domain adaptation results for dependency parsing can be achieved using a conceptually simple method that learns domain-invariant word representations. Lacking labeled resources, dependency parsing for low-resource domains has been a challenging task. Existing work considers adapting a model trained on a resource-rich domain to low-resource domains. A mainstream solution is to find a set of shared features across domains. For neural network models, word embeddings are a fundamental set of initial features. However, little work has been done investigating this simple aspect. We propose to learn domain-invariant word representations by fine-tuning pretrained word representations adversarially. Our parser achieves error reductions of 5.6% UAS, 7.9% LAS on PTB respectively, and 4.2% UAS, 3.2% LAS on Genia respectively, showing the effectiveness of domain invariant word representations for alleviating lexical bias between source and target data.