Comparing Rule-Based and Data-Driven Dependency Parsing of Learner Language

We explore the performance of two dependency parsing approaches, the rulebased WCDG approach (Foth and Menzel 2006) and the data-driven dependency parser MaltParser (Nivre et al. 2007) on texts written by language learners. We show that WCDG outperforms MaltParser in identifying the main functorargument relations, whereas MaltParser is more successful than WCDG in establishing optional, adjunct dependency relations. This can be interpreted as a tradeoff between the rich, hand-crafted lexical resources capturing obligatory argument relations in WCDG and the ability of a datadriven parser to identify optional, adjunct relations based on the linguistic and world knowledge encoded in the gold-standard training corpora.