An Empirical Etudy of Non-Lexical Extensions to Delexicalized Transfer

We propose a simple cross-language parser adaptation strategy for discriminative parsers and apply it to easy-first transition-based dependency parsing (Goldberg and Elhadad, 2010). We evaluate our parsers on the Indo-European corpora in the CoNLL-X and CoNLL 2007 shared tasks. Using the remaining languages as source data we average under-fitted weights learned from each source language and apply the resulting linear classifier to the target language. Of course some source languages and some sentences in these languages are more relevant than others for the target language in question. We therefore explore improvements of our cross-language adaptation model involving source language and instance weighting, as well as unsupervised model selection. Overall our cross-language adaptation strategies provide better results than previous strategies for direct transfer, with near-linear time parsing and much faster training times than other approaches.