Fig. 8.

Precision-Recall curves for the Evaluation with Holdout experiments. The repeat-and-average component of this training approach quite effectively smooth out the kinks observed from the Evaluation with Test Sets experiments. The differences in the gap that indicates the compatibility between the transformation rules and the observed compound structures are more visible in the Multi-Generation results. While the lower numerical values simply reflect a different measuring standard in this new approach, the expected relationships between threshold, precision and recall are preserved