The paper introduces a Dynamic Convolutional Neural Network (DCNN) for semantic modeling of sentences, which uses Dynamic $k$-Max Pooling to handle input sentences of varying lengths. The network induces a feature graph over the sentence, capturing both short and long-range relations without relying on a parse tree. The authors test the DCNN in four experiments: sentiment prediction in movie reviews, six-way question classification, and Twitter sentiment prediction using distant supervision. The DCNN outperforms other models in the first three tasks and reduces prediction error by more than 25% in the fourth task compared to the strongest baseline. The paper also discusses the background of neural sentence models, the architecture of the DCNN, and visualizes the learned feature detectors.The paper introduces a Dynamic Convolutional Neural Network (DCNN) for semantic modeling of sentences, which uses Dynamic $k$-Max Pooling to handle input sentences of varying lengths. The network induces a feature graph over the sentence, capturing both short and long-range relations without relying on a parse tree. The authors test the DCNN in four experiments: sentiment prediction in movie reviews, six-way question classification, and Twitter sentiment prediction using distant supervision. The DCNN outperforms other models in the first three tasks and reduces prediction error by more than 25% in the fourth task compared to the strongest baseline. The paper also discusses the background of neural sentence models, the architecture of the DCNN, and visualizes the learned feature detectors.