Doris Martin at SemEval-2019 Task 4: Hyperpartisan News Detection with Generic Semi-supervised Features

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 944–948 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 944 Doris Martin at SemEval-2019 Task 4: Hyperpartisan News Detection with Generic Semi-supervised Features Rodrigo Agerri IXA NLP Group, University of the Basque Country UPV/EHU rodrigo.agerri@ehu.eus Abstract In this paper we describe our participation to the Hyperpartisan News Detection shared task at SemEval 2019. Motivated by the late ar- rival of Doris Martin, we test a previously de- veloped document classification system which consists of a combination of clustering fea- tures implemented on top of some simple shal- low local features. We show how leverag- ing distributional features obtained from large in-domain unlabeled data helps to easily and quickly develop a reasonably good performing system for detecting hyperpartisan news. The system and models generated for this task are publicly available. 1 Introduction The definition of hyperpartisan according to the Hyperpartisan News Detection shared task at Se- mEval 2019 (Kiesel et al., 2019) is the following: “Given a news article text, decide whether it fol- lows a hyperpartisan argumentation, i.e., whether it exhibits blind, prejudiced, or unreasoning alle- giance to one party, faction, cause, or person”.1 Putting it simply, the task is, given a news arti- cle, to decide whether such document is hyperpar- tisan (true) or not (false). This task is related to the Stance Detection (Mohammad et al., 2016) and automatic detection of fake news (P´erez-Rosas et al., 2018) tasks, which are getting increasing attention within the Natural Language Processing community (Potthast et al., 2018). In this sense, it could be the case that hyperpartisanism is con- veyed by some elements of fake news within the article, usually with the objective of spreading pro- paganda and manipulate readers towards a partic- ular stance on a specific topic. The SemEval 2019 task 4 aims to address the problem of hyperpartisan news detection at docu- 1https://pan.webis.de/semeval19/ semeval19-web/index.html ment level, without trying to distinguish specific elements or indicators of hyperpartisanism in each article. Two sets of data were released to partici- pants. The first part (bypublisher) is annotated at publisher level. This means that if a publisher is thought to be spreading hyperpartisan news, then all its articles are annotated as hyperpartisan. The bypublisher set contains 750K articles divided in 600K documents for training and a validation set of 150K documents. The second part (byarticle) has been annotated at article level via crowdsourc- ing and consists of 645 articles for training and 628 documents for the test. The test set is hidden in TIRA (Potthast et al., 2019) and it is used for the official evaluation scores of the task. It should be noted that, unlike the byarticle test set, the byarti- cle training set was not balanced (407 false vs 238 true). We address this task using an existing document classification system, mostly due to the fact that we joined the task just a week before the final sub- mission deadline. However, and despite the lack of time …

Content truncated to 500 words