Authentic text as found in corpora cannot be described completely by a formal system, such as a set of grammar rules. As robust parsing is a prerequisite for any practical natural language processing system, there is certainly a need for techniques that go beyond merely formal approaches. Various possibilities, such as the use of simulated annealing, have been proposed recently and we have looked at their suitability for the parse process of the DLT machine translation system, which will use a large structured bilingual corpus as its main linguistic knowledge source. Our findings are that parsing is not the type of task that should be tackled solely through simulated annealing or similar stochastic optimization techniques but that a controlled application of probabilistic methods is essential for the performance of a corpus-based parser. On the basis of our explorative research we have planned a number of small-scale implementations in the near future.