Figure 11. Effects of progressive exposure to 90K corpus.

Meaning and Sentence Error for generalization to new construction (y-axis) as a function of the percentage of the corpus used for training. Generalization performance increases with percentage of corpus exposure in training. With exposure to only 12.5% of the corpus, the system generalizes with meaning error <4% and sentence error <22% for sentence final learning. Exposure to 50% of the corpus allowed generalization with 1.78% meaning error and 15.01% sentence error. Note that 12% of the sentences are ambiguous and will compete, rendering the real possible minimal sentence error ∼12%. Each point is an average over 5 instances, with means and standard deviations shown.