Cost-effective machine learning strategies for shallow parsing

For natural language applications to become widespread, they must be both practical and economical. Practicality demands that systems are robust and efficient enough to handle realistic input. Economy demands that systems are inexpensive to construct and maintain. This dissertation explores syntactic analysis of natural language from the perspective of practicality and economy. We adopt a corpus-based shallow parsing approach to syntactic analysis. Shallow parsing addresses practicality by avoiding difficult attachment decisions and by employing simple, efficient algorithms. Corpus-based language learning addresses economy by applying machine learning techniques to develop language processing components.; In particular, we contributed a new memory-based algorithm for learning shallow syntax, called rote sequence learning. Our experiments demonstrate that rote sequence learning achieves comparable performance to other, more complex, shallow parsing methods. Moreover, rote sequence learning possesses a number of desirable properties, including simplicity, efficiency, and portability. To support rote sequence learning, we developed algorithms for pruning bad rules from the grammar, for incorporating arbitrary additional information into the grammar using statistical models, and for determining the best parse among all possible parses.; Rote sequence learning addresses the practicality requirement for shallow parsing. To address economy, we investigated learning strategies that allow the machine learner to manipulate its training setting. The goal of these strategies is to reduce the cost of training by reducing the number of examples needed and/or by reducing the cost of assembling the examples. In particular, active learning allows the learner to select training examples and ask the human teacher for answers, and weakly supervised learning allows the learner to guess at answers to some of the examples on its own. In experiments with these two strategies, we discovered interesting behaviors of each. Finally, we contributed a new learning strategy, cooperative learning, that combines the best aspects of active and weakly supervised learning.