Improving Low-Resource Cross-lingual Parsing with Expected Statistic Regularization

Abstract We present Expected Statistic Regulariza tion (ESR), a novel regularization technique that utilizes low-order multi-task structural statistics to shape model distributions for semi- supervised learning on low-resource datasets. study ESR in the context of cross-lingual transfer syntactic analysis (POS tagging and labeled dependency parsing) several classes statistic functions bear behavior. Experimentally, we evaluate proposed with unsupervised 5 diverse target languages show all statistics, when estimated accurately, yield improvements both POS LAS, best improving by +7.0 LAS +8.5 average. also semi-supervised curve experiments provides significant gains over strong cross-lingual-transfer-plus-fine-tuning baselines modest amounts label data. These results indicate is promising complementary approach model-transfer approaches parsing.1