We present a new machine learning framework for multi-lingual dependency parsing. The framework uses a linear, pipeline based, bottom-up parsing algorithm, with a look ahead local search that serves to make the local predictions more robust. As shown, the performance of the first generation of this algorithm is promising.