Traditional Information Retrieval (IR) approaches assume independence between words, which is not true in reality. There have been many attempts trying to incorporate dependencies between words. Typically, a dependence model is constructed and combined with the traditional model, usually through fixed weights. However, dependencies between words vary in both strength and utility for IR. For example, the dependency in “black Friday” is more important than that in “road constructions” and considering the former can lead to much larger improvement in retrieval effectiveness than the latter. In this thesis, we propose several approaches to cope with various dependencies between different terms. In particular, we propose a method based on Dempster-Shafer theory of evidence to combine dependencies and terms, and another discriminative method to weigh different types of dependency. These methods are used for both monolingual IR and cross-language IR between English and Chinese. Significant improvements are obtained on several standard test collections.