Lexical Categories for Improved Parsing of Web Data
Lilja Øvrelid, Arne Skjærholt · International Conference on Computational Linguistics · 2012
We investigate the use of features expressing lexical generalizations over word forms when parsing web data and experiment with a range of web text samples, taken from the Ontonotes corpus, as well as the web 2.0 data sets described in Foster et al. (2011b). We obtain significant improvements for a standard data-driven dependency parser when incorporating features expressing these lexical categories, and in fact find that we may dispense with word form features altogether and still observe the same levels of improvement.