Discourse Function & Syntactic Form in Natural Language Generation

Users of natural languages have many word orders with which to encode the same truth-conditional meaning. They choose contextually appropriate strings from these many ways with little conscious effort and with effective communicative results. Previous computational models of when English speakers produce non-canonical word orders, like topicalization, left-dislocation, and clefts, fail-either by overgenerating these statistically rare forms or by undergenerating. The primary goal of this book is to present a better model of when speakers choose to produce certain non-canonical word orders by incorporating the effects of discourse context and speaker goals on syntactic choice. The theoretical model is then used as a basis for building a probabilistic classifier that can select the most human-like word order based on the surrounding discourse context. The model of discourse context used is a methodological advance both from a theoretical and an engineering perspective. It is built up from individual linguisticfeatures, ones more easily and reliably annotated than the direct annotation of a discourse or rhetorical structure for a text. This book makes extensive use of previously unexamined naturally occurring corpus data of non-canonical word order in English, both to illustrate the points of the theoretical model and to train the statistical model.