In the past few years, as automatic speech recognition has improved, we have seen a dramartic increase in the number and quality of research and commercial spoken dialogue systems. But systems which are better able to handle speech input must also be able to produce better speech output. I argue that approaches from the field of natural language generation (NLG) should be incorporated into dialog systems, and that, furthermore, methods for automatically training the natural language generator are needed to support more flexible and customized dialogs with human users. In this talk, I will focus on our recent work on automatically training the sentence planning module of a spoken language generator by learning ranking rules. The sentence planner, given the oputput of a text planner in form of communicative goals, chooses linguistic resources for realizing these goals. I present a new sentence planner SPoT, and a methodology for automatically training it using feedback provided by human judges. We show that with our training method, SPoT learns to select a sentence plan that scores within 5% on average of the sentence plans rated highest by human expert judges. Then I discuss our subject-based evaluation of SPoT, which provides independent validation for the quality of the learned rules. I conclude with a discussion of the types of rules that are needed in sentence planning, and why ranking rules may provide just the right kind of representation of the relevant knowledge, and why machine learning may be just the right way to model the acquisition of such knowledge.
Joint work with Marilyn Walker.
For the colloquium series schedule, see the UMD Computational Linguistics Colloquium Series web page at http://umiacs.umd.edu/~resnik/cl_colloquium/. If you are interested in meeting with the speaker, please contact Philip Resnik (resnik@umiacs.umd.edu).