Skip to main content
. Author manuscript; available in PMC: 2019 Oct 1.
Published in final edited form as: IEEE Trans Knowl Data Eng. 2018 Mar 5;30(10):1825–1837. doi: 10.1109/TKDE.2018.2812203

Fig. 6.

Fig. 6.

AUC curves of four variants when we have enough positive labels in the positive pool EP. Overall, human annotations lead to better results because they are more clean. However, similar trends between EPEN and DPEN show that the positive pool generated from knowledge bases has reasonable quality; the similar trends between EPEN and EPDN proves that our proposed robust positive-only distant training method works well. DPDN is the worst in this case but it has a great potential to be better as the size of positive pool grows.