Anytime Chart Realization with CCG

We describe an anytime chart realization algorithm for Combinatory Categorial Grammar (CCG) that essentially interleaves the packing and unpacking stages employed in previous hybrid symbolicstatistical approaches. With this algorithm, it becomes possible to use statistical language models both for improving the quality of the output text and for increasing the speed with which it is produced. In a case study, we describe an example-driven method of developing ngram models for use in a dialogue system, and show that with these models, the algorithm is capable of cutting realization times—and their standard deviations—in half, compared to the two-stage approach, while also achieving very high quality.