Bilingual response generation using semi-automatically-induced templates for a mixed-initiative dialog system

Abstract

We have previously developed a framework for natural language response generation for mixed-initiative dialogs in the CUHK Restaurants domain. This paper investigates the use of semi-automatic technique for response templates generation. We adopt a semi-automatic approach for grammar induction to capture the language structures of responses from unannotated corpora. We wish to use this approach to induce a set of grammars from our response data. The induced grammars are coupled with a parser to produce response templates in a semi-automatic way. Our response data consists of 2349 waiter responses. It is used as the training corpus for grammar induction. Unsupervised grammar induction is first performed, followed by using the learned grammars as prior knowledge for seeding the clustering process. Results show that the semi-automatically-induced response templates cover more than 50% of the hand-designed templates in template coverage and provide more realization options. Performance evaluation indicates that the task completion rate is at least 90%, and most of the Grice's maxims as well as the overall user satisfaction scored at 3.5 points or above.