Multi-source Domain Adaptation for Dependency Parsing via In-depth Feature Transfer

Dependency parsing endeavours to extract both syntactic and semantic knowledge of the input sentence via a dependency tree. Recently, supervised dependency parsers have achieved significant improvements thanks to the strong representation of pre-trained language models. However, the parsing accuracy drops dramatically when the model is trained on multiple out-of-domain training datasets. The key to addressing this problem is to learn the commonalities and differences between different domains. Although the widely used sharedprivate model attempts to capture domain-invariant and domainspecific feature representations by separated encoders, the two representations may interfere with each other, and the in-depth relationship between different domain features may be ignored. In this work, we explore to emphasize the useful domain-specific features and filter out the harmful ones in explicit and implicit aspects via different feature transfer strategies. Simultaneously, we leverage adversarial learning to capture more effective domain-invariant features that belong to both source and target domains. Experimental results show that our proposed model can improve the parsing accuracy of several strong baseline models significantly. The thorough analysis helps us to gain more insights into the knowledge transfer process between different modules.