Constituent Parsing by Classification

We present an approach to constituent parsing, which is driven by classifiers induced to minimize a single regularized objective. It is the first discriminatively-trained constituent parser to surpass the Collins (2003) parser without using a generative model. Our primary contribution is simplifying the human effort required for feature engineering. Our model can incorporate arbitrary features of the input and parse state. Feature selection and feature construction occur automatically, as part of learning. We define a set of fine-grained atomic features, and let the learner induce informative compound features. Our learning approach includes several novel approximations and optimizations which improve the efficiency of discriminative training. We introduce greedy completion, a new agenda-driven search strategy designed to find low-cost solutions given a limit on search effort. The inference evaluation function was learned accurately enough to guide the deterministic parsers to the optimal parse reasonably quickly without pruning, and thus without search errors. Experiments demonstrate the flexibility of our approach, which has also been applied to machine translation (Wellington et al., 2006; Turian et al., 2007).