Iterated dilated convolutional neural networks for word segmentation

The latest development of neural word segmentation is governed by bi-directional Long Short-Term Memory Networks (Bi-LSTMs) that utilize Recurrent Neural Networks (RNNs) as standard sequence tagging models, resulting in expressive and accurate performance on large-scale dataset. However, RNNs are not adapted to fully exploit the parallelism capability of Graphics Processing Unit (GPU), limiting their computational efficiency in both learning and inferring phases. This paper proposes a novel approach adopting Iterated Dilated Convolutional Neural Networks (ID-CNNs) to supersede Bi-LSTMs for faster computation while retaining accuracy. Our implementation has achieved state-of-the-art result on SIGHAN Bakeoff 2005 datasets. Extensive experiments showed that our approach with ID-CNNs enables 3× training time speedups with no accuracy loss, achieving better accuracy compared to the prevailing Bi-LSTMs. Source code and corpora of this paper have been made publicly available on GitHub.