Everything Is All It Takes: A Multipronged Strategy for Zero-Shot Cross-Lingual Information Extraction

Zero-shot cross-lingual information extraction (IE) describes the\nconstruction of an IE model for some target language, given existing\nannotations exclusively in some other language, typically English. While the\nadvance of pretrained multilingual encoders suggests an easy optimism of "train\non English, run on any language", we find through a thorough exploration and\nextension of techniques that a combination of approaches, both new and old,\nleads to better performance than any one cross-lingual strategy in particular.\nWe explore techniques including data projection and self-training, and how\ndifferent pretrained encoders impact them. We use English-to-Arabic IE as our\ninitial example, demonstrating strong performance in this setting for event\nextraction, named entity recognition, part-of-speech tagging, and dependency\nparsing. We then apply data projection and self-training to three tasks across\neight target languages. Because no single set of techniques performs the best\nacross all tasks, we encourage practitioners to explore various configurations\nof the techniques described in this work when seeking to improve on zero-shot\ntraining.\n