Team Peter Brinkmann at SemEval-2019 Task 4: Detecting Biased News Articles Using Convolutional Neural Networks

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 1032–1036 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 1032 Team Peter Brinkmann at SemEval-2019 Task 4: Detecting Biased News Articles Using Convolutional Neural Networks Michael F¨arber University of Freiburg, Germany michael.faerber@cs.uni-freiburg.de Agon Qurdina University of Prishtina, Kosovo agon.qurdina@studentet.uni-pr.edu Lule Ahmedi University of Prishtina, Kosovo lule.ahmedi@uni-pr.edu Abstract In this paper, we present an approach for clas- sifying news articles as biased (i.e., hyperpar- tisan) or unbiased, based on a convolutional neural network. We experiment with various embedding methods (pretrained and trained on the training dataset) and variations of the convolutional neural network architecture and compare the results. When evaluating our best performing approach on the actual test data set of the SemEval 2019 Task 4, we obtained relatively low precision and accuracy values, while gaining the highest recall rate among all 42 participating teams. 1 Introduction Hyperpartisan news detection describes the task of given a news article text, decide whether it fol- lows a hyperpartisan argumentation, i.e., whether it exhibits blind, prejudiced, or unreasoning al- legiance to one party, faction, cause, or person (Kiesel et al., 2019). In recent years, hyperpartisan news detection, which we consider synonymous to news bias detection, has attracted the interest of researchers and various approaches for news bias detection have been developed (Recasens et al., 2013; Baumer et al., 2015; Baly et al., 2018). However, the definition of bias and the task set-up of identifying biased news articles differs from au- thors to authors. For instance, authors might con- sider the bias in terms of the writing style, while others might consider it in relation to fact selec- tion (Hamborg et al., 2018). In this paper, we use the definition and data set of SemEval 2019 Task 4 (Kiesel et al., 2019), which deliberately uses the generic definition outlined at the begin- ning. Note that news bias detection differs from related tasks such as opinion finding (Ounis et al., 2006; Macdonald et al., 2007), sentiment analy- sis, fake news detection (Potthast et al., 2018), claim assessment (Popat et al., 2016), argumenta- tion mining on news articles (Palau and Moens, 2009), and personality detection based on texts (Mairesse et al., 2007). From a technical perspective, in recent years deep learning techniques have outperformed tra- ditional methods concerning various NLP tasks. This also applies to news article classification tasks (Ounis et al., 2006; Macdonald et al., 2007). Indeed, in the SemEval Twitter sentiment analysis competition in 2015, 2016, and 2017 (Rosenthal et al., 2015; Nakov et al., 2016; Rosenthal et al., 2017), among the most popular (and apparently effective) deep learning techniques were convolu- tional neural networks (CNNs). Thus, we decided to build a hyperpartisan news classifier based on a CNN. Next to our basic model, we also develop and evaluate variations of our model. 2 Approach In the following, we outline our approach for news bias detection.1 2.1 Preprocessing Given a set of news articles as input, we prepro- cessed them along …

Content truncated to 500 words