Orwellian-times at SemEval-2019 Task 4: A Stylistic and Content-based Classifier

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 976–980 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 976 Orwellian-times at SemEval-2019 Task 4: A Stylistic and Content-based Classifier J¨urgen Knauth Institute of Computer Science University of Goettingen jknauth@uni-goettingen.de Abstract While fake news detection received quite a bit of attention in recent years, hyperparti- san news detection is still an underresearched topic. This paper presents our work towards building a classification system for hyperpar- tisan news detection in the context of the Se- mEval2019 shared task 4. We experiment with two different approaches - a more stylistic one, and a more content related one - achieving av- erage results. 1 Introduction Recent years have seen a noticeable change in the political discourse: Political polarization has in- creased and political opinions have become more hyperpartisan (Doherty, 2017). This affects the media, especially news media and is therefore a topic of considerable interest for science and soci- ety. We present an approach for the detection of high polarization and hyperpartisan news articles. Our approach addresses stylistic and content related features. The latter are implemented by identify- ing n-grams that are typical for either a hyperpar- tisan or a more balanced perspective. 1.1 The Task The goal of the SemEval2019 Hyperpartisan News Detection Task (Kiesel et al., 2019) is to build a system capable of classifying arbitrary articles ei- ther as non-hyperpartisan or hyperpartisan. 1.2 The Dataset For building a classification system the organiz- ers of the task provided several data sets extracted from different American news sites: a) a set of 600.000 articles for training (classifica- tion: by publisher’s general orientation) b) a set of 150.000 articles for validation (classifi- cation: by publisher’s general orientation) c) 645 training articles (classified individually by humans using a crowd sourcing approach) d) validation set (classified by publisher, un- known size as this data set has been hidden dur- ing the task) e) validation set (classified individually, unknown size as this data set has been hidden during the task) All data - the articles themselves as well as the ground truth data - is provided in an propri- etary but simple and well parsable XML format defined by the task owners. The individual data records includes a globally unique ID, the title, the source URL and publication time. Addition- ally the ground truth data for a) and b) contains information about a left-right bias of the publisher in general. 2 Related Work Not so much research has been done in regard to hyperpartisan news detection. Other work is primarily addressing related fields such as ideol- ogy detection, fake news detection. For exam- ple (Hutto et al., 2015; Hamborg et al., 2018) ad- dresses identification and quantification of media bias. (Iyyer et al., 2014) addresses political ide- ology detection using recursive neural networks. (Rashkin et al., 2017) is analyzing language in fake news for automated political fact-checking. One interesting work directly targeting hyperpar- tisan news is the work of (Potthast et al., …

Content truncated to 500 words