Abstract
We study cross-lingual sequence tagging with little or no labeled data in thetarget language. Adversarial training has previously been shown to be effectivefor training cross-lingual sentence classifiers. However, it is not clear iflanguage-agnostic representations enforced by an adversarial languagediscriminator will also enable effective transfer for token-level predictiontasks. Therefore, we experiment with different types of adversarial training ontwo tasks: dependency parsing and sentence compression. We show thatadversarial training consistently leads to improved cross-lingual performanceon each task compared to a conventionally trained baseline.
Quick Read (beta)
loading the full paper ...