Relation classification using revised convolutional neural networks
Bo Li, Xiang Hui Zhao, Shuai Wang, Weihong Lin, Weidong Xiao · 2017
Relation classification plays an important part in the structuralization of text data via information extraction. Lately, neural network-based methods have been applied to relation classification, which use neural networks to encode and extract features from the input text. As convolution neural network based method can extract high-level features through convolution filters, it achieves competitive performance with other complex-structured networks by only using a standard convolution layer, a pooling layer, and a regression layer. However, it failed to capture the hierarchical and syntax information of sentence. Inspired by this, We introduce the hierarchical convolutional layers and dependency embedding to the CNN based methods. The hierarchical convolution layers capture the detail feature and high-level hierarchical features and concatenate these features as the sentence representation. The dependency embeddings help CNN capture the dependency structure in the window size, which improve the classification results. Experiments verify that the revised relation classification method provide state-of-the-art performance, even without additional artificial features.