Multi-source domain adaptation for dependency parsing via domain-aware feature generation

Abstract

With deep representation learning advances, supervised dependency parsing has achieved a notable enhancement. However, when the training data is drawn from various predefined out-domains, the parsing performance drops sharply due to the domain distribution shift. The key to addressing this problem is to model the associations and differences between multiple source and target domains. In this work, we propose an innovative domain-aware adversarial and parameter generation network for multi-source cross-domain dependency parsing where a domain-aware parameter generation network is used for identifying domain-specific features and an adversarial network is used for learning domain-invariant ones. Experiments on the benchmark datasets reveal that our model outperforms strong BERT-enhanced baselines by 2 points in the average labeled attachment score (LAS). Detailed analysis of various domain representation strategies shows that our proposed distributed domain embedding can accurately capture domain relevance, which motivates the domain-aware parameter generation network to emphasize useful domain-specific representations and disregard unnecessary or even harmful ones. Additionally, extensive comparison experiments show deeper insights on the contributions of the two components.