Weakly Supervised POS Taggers Perform Poorly on Truly Low-Resource Languages

Part-of-speech (POS) taggers for low-resource languages which are exclusively\nbased on various forms of weak supervision - e.g., cross-lingual transfer,\ntype-level supervision, or a combination thereof - have been reported to\nperform almost as well as supervised ones. However, weakly supervised POS\ntaggers are commonly only evaluated on languages that are very different from\ntruly low-resource languages, and the taggers use sources of information, like\nhigh-coverage and almost error-free dictionaries, which are likely not\navailable for resource-poor languages. We train and evaluate state-of-the-art\nweakly supervised POS taggers for a typologically diverse set of 15 truly\nlow-resource languages. On these languages, given a realistic amount of\nresources, even our best model gets only less than half of the words right. Our\nresults highlight the need for new and different approaches to POS tagging for\ntruly low-resource languages.\n