Abstract: While research on the domain adaptation task in neural machine translation has become popular recently, there exists no agreement on what constituents a domain, and most previous studies only focus on coarse-grained domain adaptation and their methods cannot be generalized if the domain size is large. In this work, we argue the necessity to study a fine-grained domain adaptation problem. We build a new multilingual dataset from web sources that focus on fine-grained domains and inter-domain attributes and relationships. We also propose a simple but effective adaptation method to incorporate domain knowledge leveraging models in information networks.
0 Replies
Loading