A Convolutional Neural Network for Modelling Sentences

摘要: The ability to accurately represent sentences is central language understanding. We describe a convolutional architecture dubbed the Dynamic Convolutional Neural Network (DCNN) that we adopt for semantic modelling of sentences. network uses k-Max Pooling, global pooling operation over linear sequences. handles input varying length and induces feature graph sentence capable explicitly capturing short long-range relations. does not rely on parse tree easily applicable any language. test DCNN in four experiments: small scale binary multi-class sentiment prediction, six-way question classification Twitter prediction by distant supervision. achieves excellent performance first three tasks greater than 25% error reduction last task with respect strongest baseline.