Files
Abstract
Pipeline systems, widely used in natural language processing and other fields, process data through sequential stages to improve efficiency, manage search complexity, and reuse components. Despite their prevalence, little systematic work has examined how these systems function or how best to improve them. This dissertation defines a formal framework based on shared properties of pipelines across domains such as parsing, speech recognition, machine translation, and image classification. It introduces quantitative metrics describing constraint characteristics and demonstrates that commonly used measures of constraint quality are poor predictors of performance. The findings show that systematically modifying constraint properties can meaningfully affect pipeline performance, offering broadly applicable guidance for analyzing and improving pipeline systems.