Rouletabille at SemEval-2019 Task 4: Neural Network Baseline for Identification of Hyperpartisan Publishers

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 981–984 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 981 Rouletabille at SemEval-2019 Task 4: Neural Network Baseline for Identification of Hyperpartisan Publishers Jose G. Moreno and Yoann Pitarch and Karen Pinel-Sauvagnat and Gilles Hubert IRIT / University of Toulouse France {jose.moreno, yoann.pitarch, karen.sauvagnat, gilles.hubert}@irit.fr Abstract This paper describes the Rouletabille partic- ipation to the Hyperpartisan News Detection task. We propose the use of different text clas- sification methods for this task. Preliminary experiments using a similar collection used in Potthast et al. (2018) show that neural-based classification methods reach state-of-the art re- sults. Our final submission is composed of a unique run that ranks among all runs at 3/49 position for the by-publisher test dataset and 43/96 for the by-article test dataset in terms of Accuracy. 1 Introduction Printed press have been in the last decades the main way to access to news in written format. This tendency is changing with the appearance of online channels but usually the main factors of the journalistic content generation are still there: events, journalists, and editors. One of the prob- lems of the generation of this content is the influ- ence of each factor in the veracity of the generated content. Two main factors may influence the final view of an article: writer’s preferences and affilia- tion of the editor house. Identifying partisan preferences in news, based only on text content, has been shown to be a chal- lenging task (Potthast et al., 2018). This problem requires to identify if a news article was written in such a way that it includes an overrated apprecia- tion of one of the participants in the news (a polit- ical party, a person, a company, etc.). Despite the fact that sharply polarized documents are not nec- essarily fake, it is an early problem to solve for the identification of fake content. A recent paper (Pot- thast et al., 2018) claims that stylometric features are a key factor to tackle this task. In this paper, we present the description of our participation to the Hyperpartisan classification Classifier Hyperpartisan Mainstream Publisher A Publisher B Figure 1: Publisher-based pipeline performed in train- ing phase. During testing, different publishers were used and labels were unknown. task at SemEval-2019 (Kiesel et al., 2019). This task was composed of two subtasks, the first con- sist to identify hyperpartisan bias in documents classified by its individual content (bias of the writer or by-article category) and the second by the editorial house that published the article (bias of the editorial house or by-publisher category) as depicted in Figure 11. To address this problem, we experimented with well-known models based on deep learning (Honnibal and Montani, 2017; Kim, 2014). They achieve state-of-the-art results on a publicly available collection (Potthast et al., 2018), showing that neural models can effectively address the task of hyperpartisan detection with- out including stylometric features. Our final sub- mission ranked in the top-3 for the by-publisher …

Content truncated to 500 words