Abstract

We made a mistake that we tested the performance of the previous approaches (Mintz’09, MultiR’11, MIML’12 and MIML-at-least-one’12) with the dataset of NYT’10 which contains much more instances labeled by NG, but plotted the precision-recall curve of our DRMC-b and DRMC-1 approaches without considering the NG-labeled testing instances. To make up the erroneous experiments, we filter out the instances with NG labels in the NYT’10 dataset and test the instances labeled by at least one positive relation in this paper. Table 1 shows the statistics of the training and testing sets. We re-run the experiments of method comparison among Mintz’09, Mul-tiR’11, MIML’12, MIML-at-least-one’12, DRMC-b and DRMC-1, and show the results in the replaced figures (Figure 4 and Figure 6). Hopefully, our algorithms still significantly outperform the state-of-art methods. Therefore, the claims in our ACL’14 paper still hold. This errata paper can be cited as: