The interpretation of persuasive discourse.

Persuasive discourse concerns the speech intent of the speaker. In order to capture the appropriate intent in context, it is crucial to recognize the internal structure of persuasive discourse. In this thesis, a computational model, called Attitude Emergence (AE), is proposed to interpret persuasive discourse. The AE model postulates a primary knowledge organization called the attitude model (or A-model), which encapsulates the speaker's speech intent in a model of his mental attitudes when making an utterance. The AE computational model performs three operations: first, attitude models are constructed for each utterance to represent the utterance's intent; second, attitude models are assimilated according to rhetorical goals; and third, the rhetorical goals, which underlie the recognition of the rhetorical and coherence relations, are verified and if applicable, related A-models are enhanced and calibrated, and implicated points derived. By formulating a domain theory of persuasion which drives the rhetorical goals, the AE computational model demonstrates that previous research on Rhetorical Structure Theory (RST) ( (MT86, MT88a)) and Speech Acts (SAT) ( (Sea69, BH79)) can be unified under an integrated model of language and attitude reasoning. It overcomes the shortcomings of RST by showing how and why rhetorical (persuasive) intent can be recognized from the rhetorical structure of the discourse. Similarly, the AE model also extends SAT with a computational model in the domain of multiple utterances which dynamically interact with each other, rather than confine SAT as a theory in the domain of a single, static utterance. The AE model is implemented in a system called BUYER on top of the Soar problem-solving architecture (LNR87). It consists of twelve problem-solving modules organized in a hierarchy. BUYER has been implemented to process real-world advertisements taken from Reader's Digest. The performance of BUYER has demonstrated that the AE model is sufficient to process real-world texts in a limited domain. Furthermore, it has also demonstrated that an integrated system which incorporates the processing of lexical cohesion, coherence and rhetorical relations, and speech intents has been achieved; while previous approaches tend to treat these processes disjointly.PhDComputer Science and EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/103340/1/9308482.pdfDescription of 9308482.pdf : Restricted to UM users only