Densely connected convolutional neural networks are currently one of the best object recognition algorithms. Given the plasticity of neural networks, the DenseNet algorithm should perform similarly in NLP tasks. In its attempt to verify whether the DenseNet algorithm can yield equally impressive results on NLP tasks, this paper has modified the DenseNet algorithm and tested it on text classification. For this purpose, three differently sized datasets have each been encoded as Tf-IDf vectors and word vectors and then the DenseNet’s performance on these different feature sets was compared to more conventional methods including Naïve Bayes classifiers and other neural networks. The paper finds that DenseNets can perform on par with these algorithms but scale especially well with large datasets and semantically rich features.
↧