Utilizing Gated Convolutional Picker to Enhance Relation Extraction

Distant supervision is an useful way to automatically gather training data for relation extraction. Aiming at the question of label marking error in distant supervision process, certain solutions are proposed. Nevertheless, the existing ways rarely think about the relation when deleting noisy data, and ignore that when the relation contained in a sentence mismatched with its label, the sentence is regarded as noise. In this research, we come up with a way to improve the relation classifier by utilizing sentence picker to eliminate the noise problem. Specifically, the gating convolution unit of the sentence picker could output the relational features, which will be utilized to determine if a sentence expresses a relationship. The relation classifier uses the high-quality training data picked by the sentence picker to enhance its performance. The two parts are trained alternately, and good performance is achieved. Experimental results tell that our module could significantly enhance the performance of relation classifier and is superior to the competitive baseline method.