Team Xenophilius Lovegood at SemEval-2019 Task 4: Hyperpartisanship Classification using Convolutional Neural Networks

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 1047–1051 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 1047 Team Xenophilius Lovegood at SemEval-2019 Task 4: Hyperpartisanship Classification using Convolutional Neural Networks Albin Zehe, Lena Hettinger, Stefan Ernst, Christian Hauptmann and Andreas Hotho DMIR Group, University of Wuerzburg <surname>@informatik.uni-wuerzburg.de Abstract This paper describes our system for the Sem- Eval 2019 Task 4 on hyperpartisan news de- tection. We build on an existing deep learning approach for sentence classification based on a Convolutional Neural Network. Modifying the original model with additional layers to in- crease its expressiveness and finally building an ensemble of multiple versions of the model, we obtain an accuracy of 67.52 % and an F1 score of 73.78 % on the main test dataset. We also report on additional experiments incorpor- ating handcrafted features into the CNN and using it as a feature extractor for a linear SVM. 1 Introduction The goal of SemEval 2019 Task 4 is to determine whether a news article blindly follows a political argumentation or not, which is referred to as ”hy- perpartisan news” (Kiesel et al., 2019). Instead of predicting the exact political orientation, it focuses on whether an article is hyperpartisan in any way. This is a very topical issue since news are easily able to reach millions of people over the internet, and in recent years have been excessively used to influence the population, for example regarding elections. Specifically, one sided media coverage influences a lot of readers without their knowledge, demonstrating the necessity of automated detection of hyperpartisan news. Approach In this work, we make use of deep learning models to address this task. We decided to adapt the sentence CNN proposed by Kim (2014), as it has been shown to be a strong baseline for text classification tasks. Since the model was originally designed for the classification of sentences, we had to make some modifications in order to deal with the longer texts provided in this task. While the very shallow model originally proposed by Kim (2014) is enough to adequately represent sentences, we found that it is not expressive enough to model entire news articles. Thus, we added a second con- volutional layer and a batch normalization layer (Ioffe and Szegedy, 2015) to the model. Addition- ally, we specified a maximum length for articles, after which they are cut off. These modifications will be described in more detail in Section 3. We also experiment with including some hand-crafted features into the model in an attempt to improve the performance. Finally, we build an ensemble of multiple models to obtain our final results. 2 Feature Extraction In order to train our models, we need to represent the input texts in some machine readable form. The data provided in the task contains multiple kinds of information that we use in different parts of our sys- tem, namely the CNN model and the hand-crafted features extracted to provide additional informa- tion. 2.1 CNN Model On …

Content truncated to 500 words