Figure 4.

Aggregated learning curves of 24 ambiguous words in the VUH corpus. (A) Interactive learning algorithms in comparison, including the best- and worst-case scenarios of “informed learning”. To achieve 90% accuracy, “random sampling” required more than 50 instance labels, “active learning” required 31 instance labels, “ReQ-ReC expert” and “Informed learning” required 26 labels. (B and C) drill-down analysis of learning curves of imperfect feature labeling (highlighting) oracles, respectively. Even using imperfect feature oracles, variants of “informed learning” still significantly outperformed “active learning”, according to Wilcoxon signed rank test (see Table 3).