Date of Award

Spring 1-1-2016

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

James H. Martin

Second Advisor

Peter W. Foltz

Third Advisor

Martha S. Palmer

Fourth Advisor

Jordan Boyd-Graber

Fifth Advisor

Eliana Colunga

Abstract

Convolutional neural networks (ConvNets) have been shown to be effective at a variety of natural language processing tasks. To date, their utility for correcting errors in writing has not been investigated. Writing error correction is important for a variety of computer-based methods for the assessment of writing. In this thesis, we apply ConvNets to a number of tasks pertaining to writing errors – including non-word error detection, isolated non-word correction, context-dependent non-word correction, and context-dependent real word correction – and find them to be competitive with or superior to a number of existing approaches. On these tasks, ConvNets function as discriminative language models, so on several tasks we compare ConvNets to probabilistic language models. Non-word error detection, for instance, is usually performed with a dictionary that provides a hard, Boolean answer to a word query. We evaluate ConvNets as a soft dictionary that provides soft, probabilistic answers to word queries. Our results indicate that ConvNets perform better in this setting than traditional probabilistic language models trained with the same examples. Similarly, in context-dependent non-word error correction, high-performing systems often make use of a probabilistic language model. We evaluate ConvNets and other neural architectures on this task and find that all neural network models outperform probabilistic language models, even though the networks were trained with two orders of magnitude fewer examples.

Share

COinS