In this dissertation, we have proposed novel methods for robust parsing that integrate the flexibility of linguistically motivated lexical descriptions with the robustness of statistical techniques. Our thesis is that the computation of linguistic structure can be localized if lexical items are associated with rich descriptions (supertags) that impose complex constraints in a local context. However, increasing the complexity of descriptions makes the number of different descriptions for each lexical item much larger and hence increases the local ambiguity for a parser. This local ambiguity can be resolved by using supertag co-occurrence statistics collected from parsed corpora. We have explored these ideas in the context of Lexicalized Tree-Adjoining Grammar (LTAG) framework wherein supertag disambiguation provides a representation that is an almost parse. We have used the disambiguated supertag sequence in conjunction with a lightweight dependency analyzer to compute noun groups, verb groups, dependency linkages and even partial parses. We have shown that a trigram-based supertagger achieves an accuracy of 92.1‰ on Wall Street Journal (WSJ) texts. Furthermore, we have shown that the lightweight dependency analysis on the output of the supertagger identifies 83‰ of the dependency links accurately. We have exploited the representation of supertags with Explanation-Based Learning to improve parsing effciency. In this approach, parsing in limited domains can be modeled as a Finite-State Transduction. We have implemented such a system for the ATIS domain which improves parsing eciency by a factor of 15. We have used the supertagger in a variety of applications to provide lexical descriptions at an appropriate granularity. In an information retrieval application, we show that the supertag based system performs at higher levels of precision compared to a system based on part-of-speech tags. In an information extraction task, supertags are used in specifying extraction patterns. For language modeling applications, we view supertags as syntactically motivated class labels in a class-based language model. The distinction between recursive and non-recursive supertags is exploited in a sentence simplification application.