We develop a novel framework that helps identify potential bias in news websites to support users who are exposed to news articles with a wide variety of political leanings. We propose a polarity-based stance (PS), a vector that represents how often a website publishes articles that are positive or negative with regard to a topic. We derive PS using the GDELT database and visualize the news websites' stances. We demonstrate the utility of our framework via a case study of the 2016 US Presidential Election.
There are two types of users when it comes to their pattern of news navigation. The rst type already has particular news websites that they trust and actively use by accessing them directly for news. Such websites tend to demonstrate the same political stances or leanings as their users. As a result, the articles that they read are likely ones that already share their ideologies. The other type, those who are less politically engaged, use a news aggregation website that shows a compiled list of news articles from various sources. A key di erence in the two approaches is that the latter
Copyright c 2018 for the individual papers by the papers' authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors. type of user, because content is the primary factor in selecting articles, is exposed to news from more diverse sources, which demonstrates a wider array of political stances. Users must therefore use their own judgment to selectively digest what they read, especially for controversial topics.
Many users judge the trustworthiness of new websites based on their political bias. Hence, we propose a novel framework that represents the bias of news websites toward a particular topic as a vector. Using this framework, we then visualize stances of news websites toward a given topic. For this, we de ne a polarity-based stance, a vector that represents bias toward a particular topic of a website using the polarity of stances. This allows us to visualize the stance of news websites, guiding users for the potential bias of the articles published by the websites. We demonstrate the usefulness of our framework via the case study of 2016 US President Election using the GDELT database1. 2
We formally de ne a polarity-based stance, P!Sw, as a two-dimensional vector that denotes the stance of a website w. We rst assume that each article of the website has one of three stances: positive, negative, or neutral. We let P!Sw = [p; n] where p is the ratio of positively-stanced articles and n is the ratio of negatively-stanced articles for a particular topic. Note that the stance has been identi ed beforehand. We discuss how to use the GDELT database to derive this vector.
1https://www.gdeltproject.org The GDELT database is one of the largest news article repositories collected by the Google Jigsaw project. It is a useful resource for multifaceted analysis for news articles because it has a large amount of data and contains the metadata including the source website that are automatically extracted from various NLP algorithms for the crawled articles [YK16].
We use tone, one type of automatically generated ! metadata, to derive P Sw. Tone refers to the average attitude of the article, which is computed by the di erence between the percentage of positive and negative terms in the document[Pro15]. Calculation of polarity score based on the term matching is simple and it is better to use more sophisticated methodology [RR15]. However, due to the large numbers of the articles for analysis, it is almost impossible for the GDELT users to crawl the all text of the articles and calculate scores for them. For the case study analysis later, we use articles from the GDELT database published on the 2016 US Presidential Election during a three month period that includes voting day (see Table 1).
! We compute P Sw using the tone score provided by the GDELT database. Let d be a news article published by a news website w and t be the tone of d. We classify the document stance sd into one of three classes: positive (1), neutral (0), and negative (-1). The stance is derived from t given a threshold using the equation sd = < t < (1) 81 > <
0 > : 1 t <
t > jwj
We then de ne a polarity-based stance (P!Sw) for a website (w) using the equation ! P Sw( ) =
Pd2w (1[sd = 1]) Pd2w (1[sd = ; (2) where w is a set of articles on published by w. By plotting these stances on a graph, users can compare stances of di erent news websites.
In addition, bias can be identi ed by comparing stances of the similar topics or one with a particular topic and general topic. jwj
We demonstrate the utility of our approach via a case study of the 2016 US Presidential Election around two topics: Donald Trump and Hillary Clinton. To visualize the polarity-based stances for these topics, we estimate the set of news articles on each topic using a simple Boolean query. When an article references both Trump and Clinton, there is ambiguity about which topic is indicated by the tone. We therefore identify the set of articles that exclusively references only one of the topic to compute the polarity-based stances (see Table 2). iheart.com yahoo.com freerepublic.com
ap.org reuters.com newsviewsnreviews.com
wn.com washingtonpost.com dailymail.co.uk alltechnews.org huffingtonpost.com avauncer.com bloomberg.com washingtonexaminer.com einnews.com
sfgate.com foxnews.com contacto-latino.com
chron.com princegeorgecitizen.com iheart.com yahoo.com freerepublic.com
ap.org reuters.com newsviewsnreviews.com
wn.com washingtonpost.com dailymail.co.uk alltechnews.org huffingtonpost.com avauncer.com bloomberg.com washingtonexaminer.com einnews.com
sfgate.com foxnews.com contacto-latino.com
chron.com princegeorgecitizen.com 1 1 1 iheart.com yahoo.com freerepublic.com
ap.org reuters.com newsviewsnreviews.com
wn.com washingtonpost.com dailymail.co.uk alltechnews.org huffingtonpost.com avauncer.com bloomberg.com washingtonexaminer.com einnews.com
sfgate.com foxnews.com contacto-latino.com
chron.com princegeorgecitizen.com and negative articles ratio for Trump and Clinton in Figure 3. We let Diff( ) to be the absolute di erence between the two components of P!Sw( ). We plot Diff(T rump) and Diff(Clinton) for comparison (See Figure 3). The websites whose bias towards the two topics are the same are plotted on the line of (Diff(T rump = Diff(Clinton)). The points at the top left of the plot are the articles that are positivelystanced towards Clinton, and the ones at the bottom right are positively-stanced towards Trump. The plot helps us identify the news websites whose polaritybased stances are completely di erent between the two topics. For example, thebostonpilot.com has (0.15, 0.27) for "Trump", and (0.16, 0.65) for "Clinton" and sci-tech-today.com as (0.02, 0.90) for "Trump", and (0.41, 0.25) for "Clinton". It is important to take into account such bias when such big di erence happens. 4
In this paper, we propose a framework to visualize stances in the dimensions of polarity of news websites to identify a potential bias in the articles that are published by them. We de ne a vector named Polaritybased Stance and demonstrate the utility via a case study of 2016 U.S. Presidential Eleciton, and that the GDELT database is a useful resource for this type of analysis. As a future work, we plan to apply our framework to a variety of topics for evaluation. We observe that some topics generally have a higher positive, or negative articles than the others. We plan to study how to take this factor into account to visualize stances in an useful way.
This work was partially supported by JSPS KAKENHI Grant Number 16H01756. [Pro15] GDELT Project. The gdelt global knowledge graph (gkg) data format codebook v2.1, 2015. [RR15] Kumar Ravi and Vadlamani Ravi. A survey on opinion mining and sentiment analysis: Tasks, approaches and applications.
Knowledge-Based Systems, 89:14 { 46, 2015.