Using Social and Linguistic Information to Adapt Pretrained Representations for Political Perspective Identification

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4569–4579 August 1–6, 2021. ©2021 Association for Computational Linguistics 4569 Using Social and Linguistic Information to Adapt Pretrained Representations for Political Perspective Identification Chang Li Department of Computer Science Purdue University, West Lafayette, IN li1873@purdue.edu Dan Goldwasser Department of Computer Science Purdue University, West Lafayette, IN dgoldwas@purdue.edu Abstract Understanding the political perspective shap- ing the way events are discussed in the media is increasingly important due to the dramatic change in news distribution. With the advance in text classification models, the performance of political perspective detection is also im- proving rapidly. However, current deep learn- ing based text models often require a large amount of supervised data for training, which can be very expensive to obtain for this task. Meanwhile, models pre-trained on the general source and task (e.g. BERT) lack the ability to focus on bias-related text span. In this pa- per, we propose a novel framework that pre- trains the text model using signals from the rich social and linguistic context that is read- ily available, including entity mentions, news sharing, and frame indicators. The pre-trained models benefit from tasks related to bias detec- tion and therefore are easier to train with the bias labels. We demonstrate the effectiveness of our proposed framework by experiments on two news bias datasets. The models with pre-training achieve significant improvement in performance and are capable of identifying the text span for bias better. 1 Introduction The perspectives underlying the way information is conveyed to readers can prime them to take similar stances and shape their worldview (Gentzkow and Shapiro, 2010, 2011). Given the highly polarized coverage of news events, recognizing these per- spectives can help ensure that all points of view are represented by news aggregation services, and help avoid “information echo-chambers” in which only a single viewpoint is represented. It may also help to prevent the spread of false information online by showing people news with different perspectives. Past work studying the expression of bias in the text has focused on lexical and syntactic represen- tations of bias (Greene and Resnik, 2009; Recasens et al., 2013; Elfardy et al., 2015). Expressions of bias can include the use of the passive voice (e.g., “mistakes were made”), or references to known ideo- logical talking points (Baumer et al., 2015; Budak et al., 2016; Card et al., 2016; Field et al., 2018; Morstatter et al., 2018) (e.g., “pro-life” vs. “pro- choice”). However, bias in news media is often nuanced and very difficult to detect. Journalists of- ten strive to appear impartial and use language that does not reveal their opinions directly. Also, by their nature, news articles describing the same real- world event will share many similar details of the event, regardless of their political perspectives. In- stead, bias is often expressed through informational choices (Fan et al., 2019), which highlight different aspects of the news story and frame facts shared by all articles in different ways. For example, the fol- lowing articles …

Content truncated to 500 words