Interpreting author intentions by analyzing story modulation

Abstract

If we are to understand human intelligence, then we need to understand human story understanding competencies, including our ability to communicate. Communication can be thought of as an externalization of an inner model of the world or an attempt to shape the inner model of the world of another. To communicate effectively, humans must analyze not only what is said, but also how it is said. My goal in this work was to develop a cognitive model of how we produce a coherent argument, explain its elements, and provide a full analysis of authorial intent. In this thesis, I propose a cognitive model of Story Modulation, or how humans glean information about a communicator's intentions or attempt to shape the inner story of their audience via key characteristics of wording. The model explains how we assemble textual evidence such as passive voice, instances of harm, and use of hedging words such as alleged, to tell a coherent story of the communicator's rhetorical goals. I demonstrate this computational model with an implementation, RASHI, that recognizes and systematically highlights intentions. The implementation reads short news-like stories in simple English and identifies modulations in text that reveal the author's intent to influence three areas-sympathy, agency, and doubt. The system gathers objective evidence using a system of modular experts, interprets the evidence with culturally-specfic subjectivity models, and distills the potentially-conflicting interpretations into a short, coherent argument about the author's intentions. I argue that RASHI, as a computational model of human communication, can be used to improve discourse surrounding the media, elevate education in critical reading, facilitate political negotiations and resolutions, and help us bridge gaps across cultures by transforming stories to be more culturally appropriate.