| Miscommunication is a serious problem with large social, economic and human costs. I develop computer-based language analysis tools to help people avoid miscommunicating with one another. Rather than relying on natural language understanding technology, I adopt the strategy of improving human communication by exposing the information that goes unstated when people communicate. I allow the user of the system to determine whether this information is a possible source of misunderstanding, and if so, how it should be resolved. I present two approaches that use this strategy. The first is a text classification system that learns to identify evidence of common misconceptions in student writing. Teachers are better able to understand, and be understood by, their students when they are aware of their students' unstated assumptions. The system is trained on selections of student writing labeled with known common misconceptions. It learns rules to identify instances of these misconceptions in the writing of other students. These are then reported back to the teacher. The classification performance compares favorably to both multiple choice questions specially crafted for a similar assessment task and inter-rater reliability scores between human judges.; The second approach is the introduction of a novel natural language processing task, that of Ambiguous Word Identification (AWI). The motivation for this task is the observation that people in a narrow discipline often develop an argot in which words that are in common usage take on specialized meanings. The use of these words helps people within the discipline communicate with one another, but they act as communication barriers when attempting to communicate with the uninitiated. The purpose of the AWI task is to identify words whose meanings vary from one domain to another. Possible applications of ambiguous word identification include semi-automatic glossary construction and a document re-targeting assistant that helps modify a document for a different audience. I design an evaluation methodology for testing the performance of AWI systems. I develop several corpus-based techniques for ambiguous word identification. Using my evaluation methodology I compare these techniques and show an improvement over several baseline measures of success. |
