Document Classification of SuDer Turkish News Corpora

Word embeddings are successfully employed in various Natural Language Processing tasks, but training them requires large amount of text, which is scarce for Turkish. In this work, we collected large amounts of articles from two news websites and tags within web pages are used as labels. Obtained corpora are tested with various document classification models. Embedding based models performed better than models with the traditional TF-IDF features. A neural network that simultaneously learns the word embeddings and document classification performed the best.