Tree-based Convolution: A New Neural Architecture for Sentence Modeling

This paper proposes a new convolutional neural architecture based on tree-structures, called the tree-based convolutional neural network (TBCNN). Two variants take advantage of constituency trees and dependency trees, respectively, to model sentences. Compared with traditional "flat" convolutional neural networks (CNNs), TBCNNs explore explicitly sentences' structural information; compared with recursive neural networks, TBCNNs have shorter propagation paths, enabling more effective feature learning and extraction. We evaluated our model in two widely applied benchmarks---sentiment analysis and question classification. Our models outperformed most state-of-the-art results, including both existing neural networks and dedicated feature/rule engineering.