From the comments section: Analyzing online public discourse on the first 2020 presidential debate

Abstract

This study explores the analysis of public reactions to presidential debates using a largely untapped social media data source—YouTube comments. My findings reveal a striking consistency in topics across YouTube channels, pointing to a shared perspective among viewers that goes beyond ideological differences. These results question the assumption that online behavior mirrors traditional media preferences, suggesting that viewers may not apply ideological biases in their selection of YouTube content, instead influenced by accessibility or algorithmic recommendations. My findings also highlight the potential of leveraging social media for identifying significant voter issues and concerns. Lastly, I highlight the distinctive nature of YouTube data and the potential of researchers to utilize it, especially in light of the limited availability of Twitter data.

Keywords

Presidential debates 2020 election social media analysis content analysis public opinion selective exposure

Introduction

Partisan media is on the rise; there is an increasing number of ideologically slanted outlets and political commentators gaining influence online (Xu et al., 2020). Lazarsfeld et al. (1948) and Stemple (1961) demonstrate selective exposure in the offline environment, meaning that participants are more likely to encounter information that affirms their existing beliefs than information that challenges their beliefs. Naturally, the emergence of online forums where people can both consume information and spread information via posts, tweets, and comments raises the possibility of selective exposure within such venues. Specifically, when commenting online, people may purposefully select videos or tweets they find appealing or agree with when deciding to comment on topics. For instance, conservatives might choose to comment on videos posted by Fox News accounts because they perceive the outlet to be more agreeable. Liberals might choose to watch CNN videos expecting more comments on topics that matter to them.

This study investigates the degree to which online users exhibit this variant of selective exposure. To do so, I examine a data source not typically utilized by social media research, YouTube comments, and focus on the first presidential debate of the 2020 United States election cycle. The goal is to determine whether there are noticeable variations in dialog among viewers on YouTube channels that host these debates. I consider questions such as whether discourse differs between viewers of the Fox News YouTube channel and viewers of CNN’s YouTube channel, and whether certain topics are more prevalent in discussions among viewers of one channel over another. For instance, is there a higher prevalence of user comments about the Black Lives Matter movement under CNN compared to Fox News?

The findings of this paper indicate people do not disproportionately post comments on different topics on specific YouTube news channels. Across mainstream media channels in the dataset, topic modeling shows almost no variation in the proportion of topics found in comments. Similarly, the analysis reveals only two channels deviate from others regarding sentiment score patterns in comments. Further analysis on previous debates in 2012 and 2016 replicates and extends these findings¹.

Selective exposure in the digital world

The first presidential debate of the 2020 election was a pivotal event, unfolding against the backdrop of a global pandemic and a significant civil rights movement (Breuninger and Wilkie, 2020). Featuring candidates Donald Trump and Joe Biden, the debate was particularly tumultuous, marked by frequent interruptions and verbal exchanges that deviated from traditional debate decorum. In addition, Chris Wallace was unusually participatory in the debate for a moderator (Breuninger and Wilkie, 2020). Following the conclusion of the debate, numerous international media outlets voiced concerns over its implications for American democracy, labeling the event a “national humiliation” for the United States (Lahut, 2020). These provocative circumstances create ideal conditions in which we would expect to see increased levels of polarization in comments across different channels.

Leveraging YouTube data to analyze reactions to political content brings distinct advantages to contemporary research on selective exposure and public sentiment, especially in the context of US Presidential debates. In contrast to traditional post-event surveys where respondents rely on their potentially faulty recollection, YouTube comments capture spontaneous reactions, which not only occur in “real time” (i.e., when the respondent reacts to a specific part of the debate) but are also unprompted. Unlike survey methodologies constrained by predetermined questions, social media analysis allows researchers to uncover issues or aspects of the debates that viewers independently choose to highlight. This form of spontaneous feedback can surface concerns or points of interest that might not be captured through structured survey instruments. Additionally, evidence of comment content variations between ideologically driven channels (e.g., CNN and Fox News), compared to neutral outlets (C-SPAN), would suggest a distinction in their respective viewer bases. For instance, a higher frequency of discussions of Black Lives Matter on CNN might indicate that CNN viewers are more interested in that topic and are also more comfortable expressing themselves on a channel they perceive to be more receptive to their ideas.

Many scholars find evidence of selective exposure’s prevalence among online information seekers (An et al., 2013; Bastos et al., 2018; Cinelli et al., 2020; Ohme and Mothes, 2020; Weeks et al., 2017). Using Facebook data, An et al.’s (2013) work demonstrates that users mainly share news articles that align with their views and avoid those that conflict, with partisans being especially prone to this behavior. Additionally, Weeks et al. (2017) find that when strong partisans are incidentally exposed to opposing views, they are more likely to actively search for and share content that aligns with their political beliefs.

On the other hand, some researchers cast doubt on the prevalence of selective exposure. Feldman et al. (2013) finds that methodological choices can significantly impact individual’s choices in selective exposure experiments, and Fletcher et al. (2020) contend that the level of polarized news engagement is overestimated². Indeed, numerous studies using varied methodologies and data sources reveal that while individuals often consume news from sources matching their political views, they also regularly engage with centrist and ideologically opposing outlets (Eady et al., 2019; Guess, 2021; Wittenberg, 2023; Yang et al., 2020). Moreover, Taneja et al. (2018) and Mangold et al. (2022) find that social media’s broad, overlapping networks of users facilitate exposure to diverse news sources and viewpoints beyond users’ close connections.

In short, current research is divided when it comes to evidence of online information seeking and dispersion. The methodologies used in such literature vary widely, including experiments, traditional surveys, and social media analyses (Cinelli et al., 2020; Eady et al., 2019; Guess, 2021; Weeks et al., 2017; Wittenberg, 2023). Multiple studies have discussed the role social media plays in political discourse and proposed new methodologies utilizing social media data (Haq et al., 2020; Németh, 2023; Rodriguez-Ibanez, 2021). Many such analyses tend to focus on the platform Twitter (now X), using Natural Language Processing (NLP) to examine political polarization on Twitter (Ali et al., 2022; Belcastro et al., 2020, 2022; Takikawa and Nagayoshi, 2017; Xia et al., 2021). However, the recent unavailability of Twitter data indicates academics must now find a new data source (Fung, 2023).

Research design

To begin, I perform an analysis of the metadata of the videos in question. This approach provides a foundational understanding and adds context to comments. Furthermore, it is important to understand the engagement patterns across different broadcasting platforms, such as which channels capture the most viewer engagement. The analysis next extends to examining sentiments expressed in the comments. This involves assessing tone intensity, gauging whether comments lean more positively or negatively. Additionally, I explore the emotional spectrum within these sentiments (anger, fear, and similar emotions). The final step involves exploratory topic modeling, where I employ the Latent Dirichlet Allocation (LDA) model to discern the variation in topics across different channels (Blei et al., 2003).

While the analysis of the metadata is primarily descriptive, two hypotheses guide the subsequent content analyses grounded in recent work arguing selective exposure is an overestimated phenomenon (Eady et al., 2019; Fletcher et al., 2020; Guess, 2021; Wittenberg, 2023; Yang et al., 2020). The first hypothesis expects minimal variation in sentiment across different channels; the second hypothesis predicts little to no variation in the distribution of topics across channels with differing ideological orientations. That is, viewers tuning into the debate via Fox News’ YouTube channel do not engage in conversations on distinct topics from those who watch on CNN’s platform.

The dataset for this study includes comments extracted from six YouTube videos showing the initial 2020 presidential debate. My selection criteria are that: videos that have high viewership, videos include the debate in its entirety, and come from mainstream outlets. The second criterion avoids selection bias that might arise from videos featuring specific moments casting the candidates in a particularly favorable or unfavorable light. I establish these criteria to capture as many mainstream audience opinions as possible.

The dataset capturing audience reactions to the first presidential debate, which aired on September 30, 2020, was compiled 1 week following the broadcast. This timing ensured that the comments represented immediate public reactions during the intense period of the 2020 election, excluding any retrospective commentary from viewers who watched later. The selected debate videos originate from various mainstream sources: C-SPAN, CBS News, CNN, Fox News, NBC News, and the Wall Street Journal. Table 1 presents detailed metadata for each of the YouTube videos analyzed, including metrics such as views, likes, dislikes, and the total number of comments received. A noteworthy observation is the significant engagement these videos garnered: across the six channels, there is a total of 36, 489, 491 views and 183,471 comments.

Table 1.

2020 Presidential debate YouTube videos metadata.

Channel	Date	Title	Views	Likes	Dislikes	Comments
CBS News	9/30/20	Trump and Biden face off in chaotic first 2020 presidential debate \| FULL DEBATE	5,435,323	49,952	5589	6220
CNN	9/30/20	Replay: The first 2020 presidential debate on CNN	2,869,477	27,284	4,359	25,488
C-SPAN	9/30/20	First 2020 presidential debate between Donald Trump and Joe Biden	13,511,529	157,387	14,993	81,057
Fox News	9/30/20	Trump–Biden presidential debate moderated by Chris Wallace \| FULL	7,127,711	112,056	8,306	35,782
NBC News	9/30/20	First presidential debate of 2020 election \| NBC news NOW	3,115,382	19,200	3870	4507
Wall Street Journal	9/30/20	Full presidential debate: President Trump and Joe Biden \| WSJ	4,430,090	45,566	4,666	30,147

C-SPAN emerges as the platform with the highest level of viewer and comment engagement, followed closely by Fox News. CNN’s metadata presents an intriguing contrast: despite having the fewest views among the channels analyzed, it contains a substantial number of comments—about 25,000 in total. This suggests that while CNN might not have attracted the largest viewing audience, its viewers are especially expressive. CBS, on the other hand, records nearly 5.5 million views but only a modest 6000 comments, hinting at a less engaged audience in comparison to CNN.

Sentiment analysis

To establish whether variation exists among commenters of different YouTube channels, I perform sentiment analysis. First I employ the VADER (Valence Aware Dictionary and sEntiment Reasoner) lexicon, a dictionary and rule-based tool specifically crafted for social media text sentiment analysis³ (Jain et al., 2023). VADER has previously captured public sentiment for tweets pertaining to the 2020 US presidential election (Ali et al., 2022; Nugroho, 2021). Nugroho (2021) shows VADER as being effective in predicting actual results of the 2020 election. Unlike other lexicons, VADER is well-suited for texts like social media comments, as it is fine-tuned for sentiments expressed in short-form content (Fontes et al., 2023). Pai et al. (2022) analyzes Twitter data with VADER, and shows the lexicon is more efficient than traditional algorithms. Thus, while VADER is a simple solution for performing sentiment analysis on social media posts, in the case of YouTube and the 2020 election, it is an effective one.

Figure 1 presents the VADER sentiment analysis outcomes across channels during the 7 days surrounding the debate. For context, recall C-SPAN received an overwhelming 80,000 comments, in contrast to CBS News’s 6,000, explaining the significant difference seen along the y-axis. The primary focus is on identifying sentiment patterns across channels. Notably, for C-SPAN, CBS News, and NBC News, positive sentiments marginally surpass negative ones, with comments in the Wall Street Journal’s video having higher prevalence of positive terminology. Fox News and CNN comments exhibit a higher use of negative language, which could reflect more pronounced partisan or ideological biases associated with these channels. I performed Levene’s test for homogeneity of variance to determine whether the variance seen between Fox News and CNN with the other channels; the results for positive (0.41), negative (0.48), and neutral (0.35) sentiment scores indicate variances between scores are not significant.

Figure 1.

VADER sentiment analysis of comments across channels over time. The VADER sentiment analysis assigns positive, neutral, and negative scores to words within a text. In this case, each comment was assigned an overall compound score classifying it into one of those categories. Figure 1 shows a breakdown of the number of positive (solid line), neutral (dotted line), and negative (dashed line) comments across channels over a seven day period following the debate.

To delve deeper into the spectrum of emotional sentiment, I employ the NRC Word-Emotion Association Lexicon, which categorizes words into eight basic emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust) and two sentiments (positive and negative) (Mohammad and Turney, 2010). I analyze the text for its overall sentiment polarity and for the specific types of emotions that are being expressed within the text. The analysis with the NRC lexicon in Figure 2 reveals a notable degree of uniformity across channels. Remarkably, the distribution of emotions and sentiment across each channel has minimal variation.

Figure 2.

NRC sentiment analysis by emotion. The NRC lexicon categorizes words with emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust) and sentiments (positive and negative). Figure 2 reveals the NRC distribution of words in comments from each channel into emotions and sentiments, revealing a remarkable consistency.

Lastly, I employ LDA topic models, which conceptualizes topics as probability distributions over words (Blei et al., 2003). Within this framework, the LDA model posits that each word in a comment can be considered part of a finite set of topics; the inclusion of each word is influenced by its association with one of these topics. After careful consideration, I specified the topic parameter generated by the LDA model to be eleven⁴.

The 11 topics identified present compelling avenues for future research. Notably, the model uncovers a significant discourse around Christian religious themes, such as references to heaven and hell, good and evil, and Jesus. These subjects, raised in comments unprompted by individuals, are not typically explored by researchers in the context of political campaigns. Additionally, the differentiation of discussions on Black Lives Matter (BLM) and COVID-19 from other general issues, such as the economy and healthcare, suggests distinct conversational threads worth further investigation. Table 2 shows the results of the LDA topic model distributed across all channel comments.

Table 2.

Proportion of topics per channel.

Channel	Religious appeals	General issues	Candidates & lying	BLM/Police	Chris Wallace	Election buzz words	Democracy	COVID-19	Watch the debate	Trump debate performance	Negative debate comments
CBS	0.089	0.090	0.091	0.091	0.090	0.092	0.092	0.089	0.091	0.093	0.092
CNN	0.090	0.090	0.091	0.091	0.093	0.091	0.090	0.089	0.091	0.092	0.091
C-SPAN	0.093	0.089	0.090	0.091	0.093	0.091	0.091	0.089	0.091	0.091	0.090
Fox News	0.087	0.089	0.091	0.092	0.088	0.091	0.089	0.099	0.090	0.091	0.093
NBC	0.089	0.102	0.089	0.090	0.089	0.089	0.091	0.088	0.090	0.092	0.089
WSJ	0.092	0.090	0.091	0.092	0.091	0.091	0.091	0.089	0.091	0.091	0.091

As with the previous sentiment analyses, the core finding here is the uniformity in all topic distributions across different media channels. This uniformity suggests that topics are discussed with similar frequency across channels as diverse as Fox News and CNN. An exploration into whether changing the number of topics (e.g., from eleven to twenty or five) would affect this uniformity yielded no significant changes, indicating a persistent consistency in how topics are proportionately discussed across channels.

To further explore the extent of topic distribution uniformity across channels, the LDA model is applied separately to comments from just CNN and Fox News. The intention is to assess whether a more focused analysis would reveal more distinct topics and greater variation between these channels. The outcomes, seen in Table 3, once again indicate significant uniformity. These results align with the hypotheses outlined earlier.

Table 3.

Proportion of topics for CNN & Fox News.

Channel	General issues	Facts & lies	Moderator-Chris Wallace	US democracy	BLM/Police topic	Candidates & lying	COVID-19	Pro-trump	Candidate performance	Negative trump sentiments	General debate comments
CNN	0.091	0.092	0.087	0.092	0.091	0.090	0.092	0.091	0.091	0.091	0.093
Fox news	0.091	0.090	0.095	0.089	0.090	0.091	0.090	0.091	0.092	0.091	0.090

Discussion

Utilizing the VADER lexicon to analyze sentiment in YouTube comments, I find that the overall sentiment trajectory is similar across most channels. While Fox News and CNN comments contain more negative language than other channels, Levene’s test for homogeneity of variance shows this distinction is not significant. The NRC analysis comparably showed no distinct variation in sentiment or emotions across all of the channels. Moreover, the investigation into topic distribution via LDA modeling demonstrates a remarkable consistency in topic uniformity across channels. This suggests that despite the diverse ideological orientations of two of the channels analyzed, the discourse within the comments maintains a similar thematic focus. This finding is particularly noteworthy, as it suggests a level of consensus or shared interest among viewers that spans across different media outlets. Furthermore, this finding holds true for analyses performed on the 2012 and 2016 presidential debates.

These results suggest that when commenting online, people may simply opt for the first video that appears in their search results. Viewers may not deliberately choose to watch the debate on a platform like C-SPAN for its inherent qualities but rather select videos based on accessibility or algorithmic suggestions. These findings imply that concerns regarding “partisan echo chambers” and selective exposure may be slightly overstated. In this way, YouTube comments present a viable alternative to Twitter data, providing viewpoints potentially less subjective to bias from selective exposure.

There is much research that outlines various NLP techniques for investigating public sentiment on Twitter (Ali et al., 2022; Belcastro et al., 2020, 2022; Takikawa and Nagayoshi, 2017; Xia et al., 2021). Specifically, many analyses rely on hashtags to identify and classify groups of interest or sentiment in texts (Belcastro et al., 2020, 2022; Ding et al., 2024). Others perform community detection and social network analysis to detect political polarization (Takikawa and Nagayoshi, 2017). However, while these methods are certainly intuitive for analyzing tweets, aspects of their methodological approach are not well-suited for YouTube comments.

In Twitter, hashtags allow users to transform specific terms into searchable links that facilitate the aggregation and discovery of tweets pertaining to similar topics. This feature allows users to easily access, contribute to, and engage with content centered around common discussions. While YouTube technologically supports hashtags, their prevalence is significantly lower compared to Twitter where hashtags are a central feature. Additionally, Twitter enables users to post directly from their accounts onto the platform, whereas YouTube requires comments to be associated with specific videos, thus context dependent. The tagging of other users and the creation of social networks, while possible on YouTube, are also much less common than on Twitter. Consequently, due to platform-specific differences in user interaction and content structuring, methodologies for Twitter analysis cannot be identically employed for YouTube data.

Part of the value of this paper is tapping into this data source that researchers generally do not use. Future scholars should build on build on this approach; leveraging YouTube comments as a resource to identify the topics and issues that hold significance to voters presents a promising research avenue. Understanding the nature of online discourse—ranging from the candidates’ debate performances to the substantive issues raised—emerges as a valuable pursuit. Such an approach not only sheds light on the electorate’s priorities and concerns but also enriches the analytical depth of content analysis in political communication studies. Moreover, the future inclusion of supervised machine learning techniques in this context could significantly enhance the precision and relevance of content analysis. By applying these computational methods, researchers can more accurately categorize, quantify, and interpret the vast and varied data generated by online discussions.

Supplemental Material

Supplemental Material - From the comments section: Analyzing online public discourse on the 2020 presidential debates

Supplemental Material for From the comments section: Analyzing online public discourse on the 2020 presidential debates by John Eaton in Research & Politics

Footnotes

Acknowledgements

I would like to thank Curtis Bram, Vito D’Orazio, and Karl Ho for their feedback and encouragement. I would also like to thank the two anonymous reviewers and the editors of Research & Politics who provided constructive comments that improve my manuscript.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

John Eaton

Supplemental Material

Supplemental material for this article is available online

Notes

References

Ali

Pinto

Lawrie

, et al. (2022) A large-scale sentiment analysis of tweets pertaining to the 2020 US presidential election. Journal of Big Data 9(1): 79. DOI: 10.1186/s40537-022-00633-z.

Quercia

Crowcroft

(2013) Fragmented social media: a look into selective exposure to political news. In: Proceedings of the 22nd International Conference on World Wide Web. New York: Association for Computing Machinery, 51–52. DOI: 10.1145/2487788.2487807.

Bastos

Mercea

Baronchelli

(2018) The geographic embedding of online echo chambers: evidence from the Brexit campaign. PLoS One 13(11): e0206841. DOI: 10.1371/journal.pone.0206841.

Belcastro

Cantini

Marozzo

, et al. (2020) Learning political polarization on social media using neural networks. IEEE Access 8: 47177–47187. DOI: 10.1109/ACCESS.2020.2978950.

Belcastro

Branda

Cantini

, et al. (2022) Analyzing voter behavior on social media during the 2020 US presidential election campaign. Social Network Analysis and Mining 12(1): 83. DOI: 10.1007/s13278-022-00913-9.

Blei

Jordan

(2003) Latent dirichlet allocation. Journal of Machine Learning Research 3: 993–1022.

Breuninger

Wilkie

(2020) Vicious first debate between Trump and Biden offered little on policy, lots of conflict. CNBC. Available at: https://www.cnbc.com/2020/09/29/first-presidential-debate-highlights-trump-vs-biden-.html, In press.

Cinelli

Brugnoli

Schmidt

, et al. (2020) Selective exposure shapes the Facebook news diet. PLoS One 15(3): e0229129. DOI: 10.1371/journal.pone.0229129.

Ding

Fan

Ding

, et al. (2024) LCSEP: a large-scale Chinese dataset for social emotion prediction to online trending topics. IEEE Transactions on Computational Social Systems 11: 3362–3375. DOI: 10.1109/TCSS.2023.3334296.

10.

Eady

Nagler

Guess

, et al. (2019) How many people live in political bubbles on social media? Evidence from linked survey and twitter data. Sage Open 9(1): 215824401983270. DOI: 10.1177/2158244019832705.

11.

Feldman

Stroud

Bimber

, et al. (2013) Assessing selective exposure in experiments: the implications of different methodological choices. Communication Methods and Measures 7(3–4): 172–194. DOI: 10.1080/19312458.2013.813923.

12.

Fletcher

Cornia

Nielsen

(2020) How polarized are online and offline news audiences? A comparative analysis of twelve countries. The International Journal of Press/Politics 25(2): 169–195. DOI: 10.1177/1940161219892768.

13.

Fontes

Murcos

Carneiro

, et al. (2023) Leveraging social media as a source of mobility intelligence: an NLP-based approach. IEEE Open Journal of Intelligent Transportation Systems 4: 663–681. DOI: 10.1109/OJITS.2023.3308210.

14.

Fung

(2023) Academic researchers blast twitter’s data paywall as ‘outrageously expensive. In: CNN Business. Atlanta: CNN. https://www.cnn.com/2023/04/05/tech/academic-researchers-blast-twitter-paywall/index.html

15.

Guess

(2021) (Almost) everything in moderation: new evidence on Americans’ online media diets. American Journal of Political Science 65(4): 1007–1022. DOI: 10.1111/ajps.12589.

16.

Haq

Braud

Kwon

, et al. (2020) A survey on computational Politics. IEEE Access 8: 197379–197406. DOI: 10.1109/ACCESS.2020.3034983.

17.

Jain

Kumar

Nayyar

, et al. (2023) Explaining sentiment analysis results on social media texts through visualization. Multimedia Tools and Applications 82: 22613–22629.

18.

Lahut

(2020) The first presidential debate was an embarrassment for the US and international news outlets came down on it harshly. In Business Insider. Available at: https://www.businessinsider.com/first-debate-international-reaction-embarrassment-for-us-on-global-stage-2020-9.

19.

Lazarsfeld

Berelson

Gaudet

(1948) The People’s Choice: How the Voter Makes up His Mind in a Presidential Election. New York, NY: Columbia University Press.

20.

Mangold

Stier

Breuer

, et al. (2022) The overstated generational gap in online news use? A consolidated infrastructural perspective. New Media and Society 24(10): 2207–2226. DOI: 10.1177/1461444821989972.

21.

Mohammad

Turney

(2010) Emotions evoked by common words and phrases: using mechanical Turk to create an emotion lexicon. In: Proceedings of the NAACL-HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, June 2010. California: Association for Computational Linguistics.

22.

Németh

(2023) A scoping review on the use of natural language processing in research on political polarization: trends and research prospects. Journal of computational social science 6(1): 289–313.

23.

Nugroho

(2021) US presidential election 2020 prediction based on Twitter data using lexicon-based sentiment analysis. In: 2021 11th International Conference on Cloud Computing, Data Science & Engineering (Confluence). Noida, India: IEEE, 136–141. DOI: 10.1109/Confluence51648.2021.9377201.

24.

Ohme

Mothes

(2020) What affects first- and second-level selective exposure to ournalistic news? A social media online experiment. Journalism Studies 21(9): 1220–1242. DOI: 10.1080/1461670X.2020.1735490.

25.

Pai

Prince

Prasannakumar

(2022) Real-time twitter sentiment analytics and visualization using vader. In: 2022 2nd International Conference on Intelligent Technologies (CONIT). Piscataway: IEEE, 1–4. DOI: 10.1109/CONIT55038.2022.9848043.

26.

Rodriguez-Ibanez

Gimeno-Blanes

Cuenca-Jimenez

, et al. (2021) Sentiment analysis of political tweets from the 2019 Spanish elections. IEEE Access 9: 101847–101862. DOI: 10.1109/ACCESS.2021.3097492.

27.

Stempel

III (1961) Selectivity in readership of political news. Public Opinion Quarterly 25(3): 400–404. DOI: 10.1086/267036.

28.

Takikawa

Nagayoshi

(2017) Political polarization in social media: analysis of the “Twitter political field” in Japan. In: 2017 IEEE International Conference on Big Data (Big Data). Piscataway: IEEE, 3143–3150. DOI: 10.1109/BigData.2017.8258291.

29.

Taneja

Edgerly

(2018) Rethinking the generational gap in online news use: an infrastructural perspective. New Media and Society 20(5): 1792–1812. DOI: 10.1177/1461444817707348.

30.

Weeks

Lane

Kim

, et al. (2017) Incidental exposure, selective exposure, and political information sharing: integrating online exposure patterns and expression on social media. Journal of Computer-Mediated Communication 22(6): 363–379. DOI: 10.1111/jcc4.12199.

31.

Wittenberg

Baum

Berinsky

, et al. (2023) Media measurement matters: estimating the persuasive effects of partisan media with survey and behavioral data. The Journal of Politics 85(4): 1275–1290. DOI: 10.1086/724960.

32.

Xia

Yue

Liu

(2021) Tweet sentiment analysis of the 2020 US presidential election. In: Companion Proceedings of the Web Conference 2021. New York: Association for Computing Machinery, 367–371.

33.

Sang

Kim

(2020) What drives hyper-partisan news sharing: exploring the role of source, style, and content. Digital Journalism 8(4): 486–505. DOI: 10.1080/21670811.2020.1761264.

34.

Yang

Majó-Vázquez

Nielsen

, et al. (2020) Exposure to news grows less fragmented with an increase in mobile access. Proceedings of the National Academy of Sciences of the United States of America 117(46): 28678–28683. DOI: 10.1073/pnas.2006089117.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.10 MB