Perceptron Training for a Wide-Coverage Lexicalized-Grammar Parser

This paper investigates perceptron training for a wide-coverage CCG parser and compares the perceptron with a log-linear model. The CCG parser uses a phrase-structure parsing model and dynamic programming in the form of the Viterbi algorithm to find the highest scoring derivation. The difficulty in using the perceptron for a phrase-structure parsing model is the need for an efficient decoder. We exploit the lexicalized nature of CCG by using a finite-state supertagger to do much of the parsing work, resulting in a highly efficient decoder. The perceptron performs as well as the log-linear model; it trains in a few hours on a single machine; and it requires only a few hundred MB of RAM for practical training compared to 20 GB for the log-linear model. We also investigate the order in which the training examples are presented to the online perceptron learner, and find that order does not significantly affect the results.