Uphill from here: Sentiment patterns in videos from left- and right-wing YouTube news channels

Felix Soldner, Justin Chun-ting Ho, Mykola Makhortykh, Isabelle W.j. van der Vegt, Maximilian Mozes, Bennett KleinbergNAACL 2019

TL;DR

This paper analyzes sentiment patterns in YouTube news videos from left- and right-wing channels, identifying seven distinct sentiment trajectories. While sentiment patterns don't strongly predict video popularity, pro-right channels tend to be more popular, and negative sentiment further boosts their viewership.

What Problem Does This Paper Address?

The increasing shift of news consumption to online platforms, particularly video-based ones like YouTube, raises concerns about the formation of ideological echo chambers. Understanding how political news is communicated and received on these platforms is crucial for comprehending its impact on public discourse and potential radicalization. This paper specifically investigates the nuances of sentiment expression in political news videos from channels with differing ideological leanings, aiming to uncover how these patterns differ and if they correlate with video popularity.

Methods

The researchers employed a linguistic temporal trajectory analysis to examine sentiment patterns within YouTube news video transcripts. This involved:

  • Data Collection: Scraping transcripts from English-language news channels on YouTube identified as having a political bias (left or right-wing) by Media Bias/Fact Check.
  • Transcript Preprocessing: Cleaning transcripts by removing XML tags and merging them into a single string without punctuation.
  • Sentiment Extraction: Utilizing a "naive context" sentiment extractor that accounts for valence shifters (negators, amplifiers, adversative conjunctions) to generate sentence-level sentiment scores. These scores were then aggregated into a continuous text sentiment trajectory.
  • Temporal Standardization: Standardizing the sentiment trajectories to a narrative time from 0 to 100, allowing for comparison across videos of varying lengths.
  • Clustering: Applying an unsupervised k-means clustering method to identify recurring sentiment patterns. The optimal number of clusters was determined to be seven.
  • Statistical Analysis: Using Chi-square tests to examine the relationship between political stance and sentiment clusters, and regression models (OLS) to investigate the predictive value of sentiment patterns and political leaning on video popularity.

Datasets

The study utilized a dataset of 15,993 video transcripts scraped from eight YouTube news channels, equally divided between pro-left and pro-right political leanings. The channels were selected from the top 250 news channels on YouTube, identified by Social Blade, and further filtered for political bias using Media Bias/Fact Check. The dataset was balanced, with approximately 2000 transcripts per channel, excluding non-English transcripts and those with insufficient word counts.

Key Insights and Ideas

  • Sentiment Styles Persist: The identification of seven distinct sentiment patterns, similar to those found in previous research on YouTube vloggers, suggests that certain sentiment "styles" are robust across different content domains on the platform.
  • Political Leanings Shape Sentiment Trajectories: There's a clear association between a channel's political leaning and the sentiment patterns it employs. Specifically, "Downhill from here" (negative sentiment) is more prevalent in pro-right channels, while "Uphill from here" (positive sentiment) is more common in pro-left channels.
  • Sentiment's Weak Predictive Power for Popularity: Contrary to some expectations, sentiment patterns alone are not strong predictors of video popularity on YouTube. This suggests other factors, potentially contextual or channel-specific, play a more significant role.
  • Pro-Right Channels and Negative Sentiment Synergy: Videos from pro-right news channels tend to be more popular, and this popularity is further amplified when the overall sentiment of the video is negative. This finding supports previous research indicating a preference for negative news tones.

Key Results

  • Seven Sentiment Clusters: The k-means clustering identified seven distinct sentiment patterns, including "Rags to riches," "Riches to rags," "Downhill from here," "Uphill from here," "End on a high note," "End on a low note," and "Mood swing."
  • Political Stance Association: A statistically significant association (p < 0.001) was found between political stance and sentiment clusters. "Downhill from here" was significantly more used by right-leaning channels, while "Uphill from here" was significantly more used by left-leaning channels.
  • Popularity Prediction: Regression models showed no significant difference in adjusted popularity rating between sentiment clusters when controlling for channel. However, an OLS regression model accounting for average sentiment and political leaning found that pro-right channels were more popular, and negative sentiment increased popularity for these channels. The model explained 24% of the variance in adjusted popularity.

Limitations and Future Work

  • Transcript Quality: The study acknowledges that transcripts could have been generated manually or automatically, potentially affecting their quality and accuracy.
  • Exclusion of Audio/Visuals: The analysis was limited to text transcripts, excluding audio and visual elements of the videos, which could influence sentiment and popularity.
  • Bias Rating Accuracy: The reliance on Media Bias/Fact Check for bias ratings means the accuracy of these ratings is a potential limitation.
  • YouTube Specificity: The focus on YouTube videos means findings might not generalize to other platforms, as news organizations may adopt different strategies for different platforms.
  • Future Work: The authors suggest future research could explore the integration of audio-visual elements, investigate how specific topics are covered by different channels, and utilize more advanced temporal analysis techniques to capture the sequential nature of sentiment trajectories for more nuanced popularity prediction.

Summary updated

BibTeX
@inproceedings{soldner2019uphill,
  author = {Felix Soldner and Justin Chun-ting Ho and Mykola Makhortykh and Isabelle W.J. van der Vegt and Maximilian Mozes and Bennett Kleinberg},
  title = {{Uphill from here: Sentiment patterns in videos from left- and right-wing YouTube news channels}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
}

Loading related content...

Figure