Projective and non-projective Turkish parsing

Ruken Cakici, Jason Michael Baldridge

Abstract: We evaluate several dependency parsing models on the METU-Sabancı Turkish Treebank through two main approaches: generative probabilistic phrase-structure parsers and discriminative dependency parsers. We find that the non-projective Maximum Spanning Tree parser of McDonald et al.(2005) is the best parsing model, in part because it recovers crossed dependencies more effectively than projective algorithms. We also show that the choice of tag set is very important and that using morphological information boosts performance, especially when the parser is not given gold standard POS tags. We also discuss some improvements to the treebank itself.