POS Taggers and Dependency Parsing

A wide-coverage parser copes with the problem of the explosion of the number of combinations of sub-trees and the number of theoretically possible dependency trees, which in the majority give spurious analyses. We show that, by using a POS tagger for choosing the most probable grammatical classes of the lexical units, we can substantially improve the rate of spurious ambiguity in a categorial dependency grammar of French developed by the NLP team of LINA. The experimental results show that our models perform better than the model which do not use a POS tagger at the cost of losing some correct analyses especially when the model of the tagger is very different to the lexical model of the parser.