Social feedback amplifies emotional language in online video live chats

TL;DR

It is found that emotional expressions are up to four times more strongly driven by peer interaction than by video content, with asymmetric cross-excitation, with negative emotions frequently triggering positive ones, a pattern consistent with trolling dynamics, but not the reverse.

Abstract

A growing share of human interactions now occurs online, where the expression and perception of emotions are often amplified and distorted. Yet, the interplay between different emotions and the extent to which they are driven by external stimuli or social feedback remains poorly understood. We calibrate a multivariate Hawkes self-exciting point process to model the temporal expression of six basic emotions in YouTube Live chats. This framework captures both temporal and cross-emotional dependencies while allowing us to disentangle the influence of video content (exogenous) from peer interactions (endogenous). We find that emotional expressions are up to four times more strongly driven by peer interaction than by video content. Positivity is more contagious, spreading three times more readily, whereas negativity is more memorable, lingering nearly twice as long. Moreover, we observe asymmetric cross-excitation, with negative emotions frequently triggering positive ones, a pattern consistent with trolling dynamics, but not the reverse. These findings highlight the central role of social interaction in shaping emotional dynamics online and the risks of emotional manipulation as human-chatbot interactions become increasingly realistic.

Figures (10)

• Figure 1: (a) Screenshot of a YouTube live video on the Theranos scandal involving Elizabeth Holmes with a live chat section on the right. The video content is captured by transcripts (subtitles), which we use to proxy exogenous emotional stimuli for the live chats. The live chat section displays timestamped messages from users reacting to the video content in real-time as the live video streams. We highlight an example of a transcript labeled as sad in blue, and a live chat message labeled as angry in red. (b,c) We visualize the extraction of emotions from the live chat in the screenshot above. We plot a subset of live chat messages from the video sample that are labeled as sad (angry), indicated with diamond markers in blue (red). The shared x-axis shows the time in the video in units of minutes. The y-axis is unit-less. We label emotions non-exclusively. For instance, the sentence "Holmes is a victim of the fake news media." is labeled as both sad and angry. We assume these emotions are generated by the latent, inhomogeneous intensity defined by expression \ref{['eq:multi_hawkes_general_main']}. (d) Time-varying component of the emotion sad in the video (transcript). Blue dots annotate the arrival of transcripts that are labeled as sad, shown with a 2-second rightward shift for alignment with peaks in the signal for visual convenience (see Methods section for details). The black dashed line represents the temporal function $S^\text{sad}(t)$, capturing the presence of sad emotions within the video.
• Figure 1: Number of videos per keyword in our final data sample across 27 keywords.
• Figure 2: Visualizations of estimated parameters $\nu^{e,f}$, $\mu^{e}_0$, $\alpha^{e,f}$, and $\gamma^e$ of the intensity defined by expression \ref{['eq:multi_hawkes_general_main']} from the maximization of the log-likelihood function \ref{['eq:multi_cross_mu_main']}. The parameters are fitted ten times, each time sampling 60% (238) of the total 397 videos at random. Values and error bars are then obtained as sample mean and standard deviations across all 100 fits, respectively. Entry $\alpha^{e,f}$ of the $\alpha$ matrix represents excitation from emotion $f$ to emotion $e$ in the live chat, while $\gamma^e$ represents the characteristic time-scale over which past emotions trigger new emotions $e$. The $\nu$ matrix demonstrates how emotions in the video trigger emotions in the live chat, while $\mu_0$ illustrates the spontaneous baseline intensity of spontaneous emotion expression. Note that the $\gamma^e$ and $\mu_0^e$ columns share the same $y$-axis emotion labels as the $\alpha$ and $\nu$ matrices, respectively.
• Figure 2: (left) Distribution of the number of events per video across 397 videos. (right) Same as left but broken down per emotion.
• Figure 3: (a) For each video, we calculate the average ratio of endogenous (exogenous) intensity to the total intensity across time. Plot (a) shows the distribution of these ratios across $n = 397$ videos. We notice that emotions are predominantly triggered endogenously. (b) Same as plot (a) but for the ratio of spontaneous (video-influenced) intensity to total exogenous intensity.