As more and more treebanks, i.e. syntactically annotated
corpora, become available for a wide variety of languages, 
machine learning approaches to parsing gain interest as a 
means of developing parsers without having to repeat such 
labor-intensive and language-specific activities as grammar 
development for each new language. In this paper, we
describe two different machine learning approaches to the 
CoNLL-X shared task on multi-lingual dependency parsing. 
First, we introduce a number of baselines that generate 
left-branching, right-branching or more complex trees. 
Next, we present two systems that were submitted to the 
shared task: 1) an approach that directly predicts all 
dependency relations in a single run over the input 
sentence, and 2) a cascade of phrase recognizers. We 
find that the first approach performs best and conclude 
with a detailed error analysis of its output for two of 
the thirteen languages in the task, Dutch and Spanish.