Traditional syntactic models of parsing have been inadequate for task-driven processing of extended text, because they spend most of their time on misdirected linguistic analysis, leading to problems with both efficiency and coverage. Statistical and domain-driven processing offer compelling possibilities, but only as a complement to syntactic processing. For semantically-oriented tasks such as data extraction from text, the problem is how to combine the coverage of these "weaker" methods with the detail and accuracy of traditional lingusitic analysis. A good approach is to focus linguistic analysis on relations that directly impact the semantic results, detaching these relations from the complete constituents to which they belong. This approach results in a faster, more robust, and potentially more accura.te parser for real text.