Breaking the barrier of context-freeness: towards a linguistically adequate probabilistic dependency model of parallel texts.