A Framework for Creating Text Parsing Dialogue Systems

Dialogue-based simulations are simulations designed to teach users methods for interacting with a target population. They have a wide variety of applications including training medical students to interact with patients, teaching military personnel about local cultures before deployment, and giving students learning a foreign language a chance to practice by going through scenarios. Many dialogue-based simulations have used a multiple-choice dialogue system. While there has been recent work on free-input dialogue systems for educational simulations, most frameworks for free-input dialogue systems do not preserve the dialogue tree structure of a multiple-choice dialogue system. This thesis aims to create a framework for transitioning from a multiple-choice dialogue system to a free-input text classification system. The framework includes methods for crowdsourcing for data collection, a binary sub-category data labeling system, and a data generation algorithm. An implementation of the proposed framework in an existing military educational simulation is developed. This implementation replaces the multiple-choice dialogue system with classification models trained on datasets created using the framework. A test of the framework is performed on real user input. The results indicate that this approach may offer a viable method for building free-input dialogue systems for educational simulations.