Very Deep Convolutional Networks for Text Classification

TL;DR

This paper introduces VDCNN, a very deep convolutional neural network for text classification that operates at the character level, demonstrating improved performance over existing models on multiple benchmarks.

Contribution

First application of very deep convolutional networks to text processing, showing that increased depth improves classification performance.

Findings

Deeper models outperform shallower ones on text classification tasks.

VDCNN achieves state-of-the-art results on several benchmarks.

Character-level processing is effective for deep convolutional architectures.

Abstract

The dominant approach for many NLP tasks are recurrent neural networks, in particular LSTMs, and convolutional neural networks. However, these architectures are rather shallow in comparison to the deep convolutional networks which have pushed the state-of-the-art in computer vision. We present a new architecture (VDCNN) for text processing which operates directly at the character level and uses only small convolutions and pooling operations. We are able to show that the performance of this model increases with depth: using up to 29 convolutional layers, we report improvements over the state-of-the-art on several public text classification tasks. To the best of our knowledge, this is the first time that very deep convolutional nets have been applied to text processing.