tree: c45bb7e457baa8247eac6b846f738ee5b608702d [path history] [tgz]

Implementing CNN + Highway Network for Chinese Text Classification in MXNet

Sentiment classification forked from incubator-mxnet/cnn_text_classification/, i've implemented the Highway Networks architecture.The final train model is CNN + Highway Network structure, and this version can achieve a best dev accuracy of 94.75% with the Chinese corpus.

It is a slightly simplified implementation of Kim's Convolutional Neural Networks for Sentence Classification paper in MXNet.

Recently, I have been learning mxnet for Natural Language Processing (NLP). I followed this nice blog “Implementing a CNN for Text Classification in Tensorflow” blog post. to reimplement it by mxnet framework. Data preprocessing code and corpus are directly borrowed from original author cnn-text-classification-tf.

Performance compared to original paper

I use the same pretrained word2vec GoogleNews-vectors-negative300.bin in Kim‘s paper. However, I don’t implement L2-normalization of weights on penultimate layer, but provide a L2-normalization of gradients. Finally, I got a best dev accuracy 80.1%, close to 81% that reported in the original paper.


Please download the corpus from this repository cnn-text-classification-tf, :)

‘data/rt.vec’, this file was trained on the corpus by word2vec tool. I recommend to use GoogleNews word2vec, which could get better performance, since this corpus is small (contains about 10K sentences).

When using GoogleNews word2vec, this code loads it with gensim tools gensim.


If I were wrong in CNN implementation via mxnet, please correct me.