Publication at Faculty of Mathematics and Physics |
2016
Abstract
In this paper, we compare delexicalized transfer and minimally supervised parsing techniques on 32 different languages from Universal Dependencies treebank collection. The minimal supervision is in adding handcrafted universal grammatical rules for POS tags.
The rules are incorporated into the unsupervised dependency parser in forms of external prior probabilities. We also experiment with learning this probabilities from other treebanks.
The average attachment score of our parser is slightly lower then the delexicalized transfer parser, however, it performs better for languages from less resourced language families (non-Indo-European) and is therefore suitable for those, for which the treebanks often do not exist.