This paper presents a comparison of different methods for training conditional random fields (CRFs) for shallow parsing, showing that CRFs outperform other methods on the CoNLL-2000 task. The authors demonstrate that CRFs can achieve performance as good as any reported base noun-phrase chunking method and better than any single model. They also show that improved training methods based on modern optimization algorithms are critical in achieving these results. The paper compares various training methods and confirms previous results on shallow parsing and training methods for maximum-entropy models. The authors also describe the CRF model and its training methods, including preconditioned conjugate gradient, limited-memory quasi-Newton, and voted perceptron. They show that CRFs can be used for shallow parsing, with a second-order Markov dependency between chunk tags. The paper also discusses parameter tuning, evaluation metrics, and significance tests. The results show that CRFs outperform other methods on the CoNLL-2000 task, with the best F score being 94.07%. The paper concludes that discriminative sequence labeling models trained with general-purpose optimization methods are a simple, competitive solution to learning shallow parsers. The authors also discuss the potential of (log-)linear parsing models to supplant lexicalized PCFG models for parsing.This paper presents a comparison of different methods for training conditional random fields (CRFs) for shallow parsing, showing that CRFs outperform other methods on the CoNLL-2000 task. The authors demonstrate that CRFs can achieve performance as good as any reported base noun-phrase chunking method and better than any single model. They also show that improved training methods based on modern optimization algorithms are critical in achieving these results. The paper compares various training methods and confirms previous results on shallow parsing and training methods for maximum-entropy models. The authors also describe the CRF model and its training methods, including preconditioned conjugate gradient, limited-memory quasi-Newton, and voted perceptron. They show that CRFs can be used for shallow parsing, with a second-order Markov dependency between chunk tags. The paper also discusses parameter tuning, evaluation metrics, and significance tests. The results show that CRFs outperform other methods on the CoNLL-2000 task, with the best F score being 94.07%. The paper concludes that discriminative sequence labeling models trained with general-purpose optimization methods are a simple, competitive solution to learning shallow parsers. The authors also discuss the potential of (log-)linear parsing models to supplant lexicalized PCFG models for parsing.