Article details

Research area
Speech recognition

Location
NAACL/HLT 2012

Date
2012

Author(s)
Matt Hohensee, Emily M. Bender

Getting more from morphology in multilingual dependency parsing

Synopsis:

We propose a linguistically motivated set of features to capture morphological agreement and add them to the MSTParser dependency parser. Compared to the built-in morphological feature set, ours is both much smaller and more accurate across a sample of 20 morphologically annotated treebanks. We find increases in accuracy of up to 5.3% absolute. While some of this results from the feature set capturing information unrelated to morphology, there is still significant improvement, up to 4.6% absolute, due to the agreement model.

Read/download now