Investigation of Weakly Supervised Learning for Semantic Role Labeling

Abstract

In this paper, we investigate the possibility of the weakly supervised learning for Semantic Role Labeling. First, we attempt to achieve feature splitting which is the base constraint of co-training, and examine if co-training works for the task of Semantic Role Labeling. We also examine the possibility of self-training which uses the identical features with co-training, and compare the performance of co-training and self-training. From the experiments, we found some interesting points about Semantic Role Labeling task and the weakly supervised learning. As far as we know, this is the first experiment to apply weakly supervised learning to Semantic Role Labeling and the experimental results show that Semantic Role Labeling can be successfully done by weakly supervised learning.