Inference with classifiers: a study of structured output problems in natural language processing

摘要: A large number of problems in natural language processing (NLP) involve outputs with complex structure. Conceptually such problems, the task is to assign values multiple variables which represent several interdependent components. approach this formulate it as a two-stage process. In first stage, are assigned initial using machine learning based programs. second, an inference procedure uses outcomes stage classifiers along domain specific constraints order infer globally consistent final prediction. This dissertation introduces framework, classifiers, study problems. The framework applied two important and fundamental NLP that structured outputs, shallow parsing semantic role labeling. parsing, goal identify syntactic phrases sentences, has been found useful variety large-scale applications. Semantic labeling identifying predicate-argument structure crucial step toward deeper understanding language. both tasks, we develop state-of-the-art systems have used practice. In shown significance incorporating into way correct improve decisions stand alone classifiers. Although clear necessarily improves global coherency, there no guarantee improvement performance measured terms accuracy local predictions---the metric interest for most We better theoretic issue. Under reasonable assumption, prove sufficient condition cannot degrade respect Hamming loss. addition, provide experimental suggesting can even when conditions not fully satisfied.