Expectation has been argued to determine the processing difficulty of crossing dependencies (e.g., right-extraposed relative clause) in English (Levy et al., 2012). It has also been argued that this difficulty can be determined by memory reactivation, prediction maintenance, etc. (Husain and Vasishth, 2015). Since the verb can be strongly expected in head-final languages, the expectation-based account will predict that crossing dependencies will be easier to process in such languages and will not be affected by working memory constraints. We investigate this possibility in Hindi using a self-paced reading experiment (N=60). The experiment uses non-finite clause constructions in Hindi where a modifier of finite verb intervenes the non-finite clause in crossing dependency conditions. The results show that maintenance of a crossing dependency is contingent on the intervening material in a head-final language. If the intervening material is not compatible with the original prediction then the prediction gets revised –- incompatibility of NP=Accusative/Ergative with the original prediction (intransitive NFV) triggers a prediction revision that leads to slowdown at NFV. This revision cost can be explained by the expectation-based account; it cannot be explained by an activation-based account. The revision cost also shows that if possible the parser tries to build a non-crossing dependency, in spite of a strong context-based prediction. Results are consistent with Levy et al. (2012) –- processing crossing dependencies even in a head-final configuration is difficult and strong expectation attenuates this difficulty. However, the cost of processing crossing dependency in a head-final language reflects different aspects of the prediction processes -- prediction maintenance in certain dependencies will be contingent on whether the prediction gets revised due to the intervening material or not.