Parsing Indian Languages with MaltParser
Joakim Nivre · 2009
This paper describes the application of MaltParser, a transition-based dependency parser, to three Indian languages – Bangla, Hindi and Telugu – in the context of the NLP Tools Contest at ICON 2009. In the final evaluation, MaltParser was ranked second among the participating systems and achieved an unlabeled attachment score close to 90 % for Bangla and Hindi, and over 85 % for Telugu, while the labeled attachment score was 15–25 percentage points lower. It is likely that the high unlabeled accuracy is achieved thanks to a relatively low syntactic complexity in the data sets, while the low labeled accuracy is due to the limited amounts of training data.