On Minimizing Training Corpus for Parser Acquisition

Abstract

Many corpus-based natural language processing systems rely on using large quantities of annotated text as their training examples. Building this kind of resource is an expensive and labor-intensive project. To minimize effort spent on annotating examples that are not helpful the training process, recent research efforts have begun to apply active learning techniques to selectively choose data to be annotated. In this work, we consider selecting training examples with the tree-entropy metric. Our goal is to assess how well this selection technique can be applied for training different types of parsers. We find that tree-entropy can significantly reduce the amount of training annotation for both a history-based parser and an EM-based parser. Moreover, the examples selected for the history-based parser are also good for training the EM-based parser, suggesting that the technique is parser independent. Cross-referenced as UMIACS-TR-2001-40 Cross-referenced as LAMP-TR-073