Incremental interpretation and production across clauses

Thesis (Ph. D.)--University of Rochester. Dept. of Computer Science and Dept. of Linguistics, 2011.There is ample evidence that humans process language incrementally. Comprehenders decode the linguistic signal as it unfolds over time. Speakers start articulating as soon as parts of the intended, preverbal message have been encoded linguistically. Yet, much remains unknown about the incremental comprehension and production of a message. In particular psycholinguistic work has mostly been limited to behavioral experiments, thus relatively little is known about incremental production of spontaneous speech. In this dissertation, we use richly annotated speech corpora to investigate how speakers interpret and convey a message across clauses. This dissertation has three parts. Part I introduces two novel corpora. One of these has multiple layers of annotation that make it ideally suited to psycholinguistics and computational linguists working on incremental processing. Part II presents a case study of how the same methods can be used to develop computational models of parsing in a conversational dialog architecture. Part III presents case studies that show how computational methods combined with richly annotated corpora can be used to study incremental production in task-based communication. Part II discusses standard computational models developed in the field of Natural Language Processing. These models proceed in essentially an off-line manner. Parsing proceeds sentence-by-sentence, so that interpretation only becomes available after an entire sentence is processed. These models are incompatible with psycholinguistic findings. This part presents a proof of concept that semantic expectations can be incorporated in the probability models used in automatic parsing as the linguistic signal unfolds. Part III presents a series of studies which employ the above mentioned corpora to investigate production beyond the clausal level, an area that has so far received relatively little attention, in part due to the complexities involved in approaching the question experimentally. We explore why speakers express a message in one or two clauses. These studies show that an account based on the optimization of limited resources available to the human processor explains the observed behavior: the more complex the first part of the message, the more likely speakers use two clauses. We show that such account can be extended to incorporate information density as a metric of complexity. Taken together, the findings in this dissertation contribute to the growing literature that merges two traditions of studying language as action and as product. Information theoretic principles are used as a new tool for exploring on-line processes in spontaneous cooperative communication