Do You Feel What I Feel? Social Aspects of Emotions in Twitter Conversations

Suin Kim, JinYeong Bak, Alice Oh

Abstract: We present a computational framework for understanding the social aspects of emotions in Twitter conversations. Using unannotated data and semisupervised machine learning, we look at emotional transitions, emotional influences among the conversation partners, and patterns in the overall emotional exchanges. We find that conversational partners usually express the same emotion, which we name Emotion accommodation, but when they do not, one of the conversational partners tends to respond with a positive emotion. We also show that tweets containing sympathy, apology, and complaint are significant emotion influencers. We verify the emotion classification part of our framework by a human-annotated corpus.

Introduction

In the early 2010s, as social media became a primary mode of global communication, researchers began to investigate whether the "social" in social media extended to emotional exchange. While sentiment analysis—the binary classification of text as positive or negative—was already a mature field, it often treated tweets as isolated data points. This approach overlooked the interactive nature of human emotion: we do not just broadcast feelings; we exchange, mirror, and influence them.

In their study, "Do You Feel What I Feel? Social Aspects of Emotions in Twitter Conversations," Suin Kim, JinYeong Bak, and Alice Oh from KAIST developed a framework to map these emotional dynamics. Their work identifies how emotions transition within "chains" of replies and which specific topics act as pivots for the collective mood of a digital conversation.

Methodology

The researchers developed a multi-stage pipeline to transform raw Twitter data into a structured map of emotional influence.

Topic Discovery and Lexicon Expansion

The team used Sentence-LDA, a variation of Latent Dirichlet Allocation, to categorize tweets into 200 fine-grained topics. This allowed for a more nuanced understanding of emotional triggers—distinguishing, for example, between anger directed at technical issues versus personal grievances. To identify emotions within these topics, they started with a "seed" set of words for eight basic emotions (Joy, Anger, Fear, etc.) based on Plutchik’s wheel. They then used Pointwise Mutual Information (PMI) to expand this lexicon to include Twitter-specific slang and misspellings (e.g., "lmao," "wowww") that strongly correlated with those emotions.

The Sliding Window

To quantify influence, the study employed a three-tweet sliding window. This mechanism allowed the researchers to observe how a specific topic introduced by one user might alter the emotional state of their partner in the subsequent reply.

The sliding window framework used to track emotional influence between conversational partners.

Dataset and Evaluation

The study analyzed a massive corpus of English Twitter data from 2011, filtering for meaningful interactions. The final dataset consisted of 153,054 conversation chains containing over 871,000 tweets. To validate their computational model, the authors compared its results against a human-annotated corpus of 1,000 tweets.

Key Insights and Results

Emotion Accommodation

The most significant finding was the phenomenon of Emotion Accommodation. The data revealed that the most frequent emotional transition is a "self-transition." If a user receives a joyful tweet, they are most likely to respond with joy. This suggests that digital interlocutors subconsciously synchronize their emotional frequencies.

State transition diagrams showing (a) probabilities between specific emotions and (b) the transition between emotional and neutral states.

As shown in the transition diagrams, the probability of a "Joy" tweet being met with a "Joy" response was 0.51—the highest transition probability observed. Furthermore, the Twitter landscape is heavily skewed toward positive interactions; Joy (39.7%) and Anticipation (15.1%) were the most prevalent emotions, while Fear and Disgust were relatively rare.

The Positivity Pivot and Strategic Influencers

When users do not mirror a negative emotion, they rarely respond with a different negative one. Instead, there is a strong tendency to pivot toward positive emotions like Joy or Acceptance. The researchers identified specific "power player" topics that facilitate these shifts:

  • Positive Influencers: Topics related to greetings, sympathy, and recommendations were most effective at moving a partner from a negative to a positive state.
  • Negative Influencers: Conversely, topics involving worry and complaints were the most potent at dragging a positive conversation into a negative emotional state.

Performance and Limitations

The computational framework achieved a classification accuracy of 31.3%. While this figure may appear modest, it was double the accuracy of simple word-counting methods and significantly outperformed random baselines. For context, the researchers noted that human annotators only agreed on emotion classification approximately 54.4% of the time, highlighting the inherent ambiguity of short-form digital text.

The study acknowledges several limitations, including the potential for bias due to the high prevalence of "Joy" on the platform and the difficulty of detecting sarcasm. Future research directions include exploring temporal patterns—how long emotional states persist—and identifying "emotional anchors" versus "mirrors" among different user profiles.

Summary updated

Abstract

We present a computational framework for understanding the social aspects of emotions in Twitter conversations. Using unannotated data and semisupervised machine learning, we look at emotional transitions, emotional influences among the conversation partners, and patterns in the overall emotional exchanges. We find that conversational partners usually express the same emotion, which we name Emotion accommodation, but when they do not, one of the conversational partners tends to respond with a positive emotion. We also show that tweets containing sympathy, apology, and complaint are significant emotion influencers. We verify the emotion classification part of our framework by a human-annotated corpus.
BibTeX
@inproceedings{kim2021do,
  author = {Suin Kim and JinYeong Bak and Alice Oh},
  title = {{Do You Feel What I Feel? Social Aspects of Emotions in Twitter Conversations}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v6i1.14310},
}
Figure