Very Deep Convolutional Networks英文文献资料.pdf

Very Deep Convolutional Networks for Text Classification Alexis Conneau Holger Schwenk Yann Le Cun Facebook AI Research Facebook AI Research Facebook AI Research aconneau@ schwenk@ yann@ ¨ Loıc Barrault LIUM, University of Le Mans, France 7 loic.barrault@univ-lemans.fr 1 0 2 Abstract terest in the research community and they are sys- n tematically applied to all NLP tasks. However, a The dominant approach for many NLP while the use of (deep) neural networks in NLP J tasks are recurrent neural networks, in par- has shown very good results for many tasks, it 7 ticular LSTMs, and convolutional neural seems that they have not yet reached the level to 2 networks. However, these architectures outperform the state-of-the-art by a large margin, ] are rather shallow in comparison to the as it was observed in computer vision and speech L C deep convolutional networks which have recognition. . pushed the state-of-the-art in computer vi- Convolutional neural networks, in short Con- s c sion. We present a new architecture (VD- vNets, are very successful in computer vision. In [ CNN) for text processing which operates early approach