Cross-View Adaptation Network For Cross-Domain Relation Extraction

In relation extraction, directly adopting a model trained in the source domain to the target domain will suffer greatly performance decrease. Existing studies extract the shared features between domains in a coarse-grained way, which inevitably introduce some domain-specific features or suffer from information loss. Inspired by human beings often using different views to find connection between domains, we argue that, there exist some fine-grained features which can be shared across different views of origin data. In this paper, we proposed a cross-view adaptation network, which use adversarial method to extract shared features and introduce cross-view training to fine-turn it. Besides, we construct some novel views of input data for cross-domain relation extraction. Through experiments we demonstrated that the different views of data we construct can effectively avoid introducing some domain-specific features into unified feature space and help the model learn a fine-grained shared features of different domain. On the three different domains of ACE 2005 dataset, Our method achieved the state-of-the-art results in F1-score.