German News Article Classification: A Multichannel CNN Approach

Nowadays, more and more people are gaining interest in news and social media networks, and are also sharing their opinions freely in different languages. Such kind of activities leads to interesting topics of research that scientists are working on. Considering news, it must be classified and easily accessible by the users for the information of their interest. In comparison with traditional machine learning techniques, deep learning approaches have achieved surpassing results on natural language processing tasks. Convolutions neural networks (CNNs) have shown promising performance, which extracts n-grams as features to represent the input. In this work, we build a multi-channel CNN for German news article classification. The model can classify different categories of news articles with an accuracy of 99.2% on training and 81.4% on the test dataset. We also perform a comparative study with single-channel CNN and have found that the multi-channel approach outperforms the single-channel by +6.3% absolute on the test set.