Cross Domain Few-Shot Learning via Meta Adversarial Training

Previous research on few-shot relation classification(RC) only considers the situation that the datasets for meta-training and meta-testing are from the same domain until the presentation of FewRel 2.0 (Gao et al., 2019b) dataset, whose metatraining and meta-testing differ vastly from each other: the meta-training data derives from Wikidata, a comprehensive and non-professional dataset, while meta-testing data is from Pubmed, a medical dataset greatly different from Wikidata in morphology and syntax, which leads to a novel cross-domain few-shot learning problem. Some previous works adopt adversarial training to solve the cross-domain problems, like Prototypical-ADV model (Gao et al., 2019b), which introduces a adversarial training method with a number of unlabeled target-domain data to make the encoded target-domain and source-domain instances more similar. Moreover, on the basis of it, (Cong et al., 2020) further uses the unlabeled target-domain data. The unlabeled target-domain data gets pseudo-labels generated by cluster miner, and is used for meta-training. In these previous works, they use additional unlabeled target-domain (meta-testing) instances to do the adversarial training in the meta-training process.