Abstract:

I have developed an implementation to address a natural language understanding task called "Dialogue Act Tagging" with a machine learning method called "Transformation-Based Learning". It was necessary for me to 1) develop a Monte Carlo approach to make the method tractable for my task, 2) adapt a committee method to compute useful confidence measures for the system's tags, and 3) investigate strategies for automatically selecting "dialogue act cues", the most useful words or phrases for computing dialogue acts in a given domain. With these contributions, my implementation handles the dialogue act tagging task as effectively as any other reported system.