Exploiting Community Emotion for Microblog Event Detection

Gaoyan Ou, Wei Chen, Tengjiao Wang, Zhongyu Wei, Binyang Li, Dongqing Yang, Kam-Fai Wong

Introduction

Traditional microblog event detection primarily relies on identifying "bursty" keywords, such as a sudden spike in the word "earthquake." While effective for major global occurrences, this approach often overlooks the rich emotional context of user reactions and fails to detect niche events relevant only to specific sub-communities (e.g., legal professionals or football fans). Furthermore, global aggregation can drown out significant local signals. This paper introduces a novel framework that leverages community-specific emotional dynamics to detect real-world events, while mitigating the impact of "emotion spam" through social authority weighting.

Methodology

The proposed system follows a three-stage pipeline to transform raw microblog data into actionable event detections, focusing on the "vibe" of specific online communities rather than just global keyword counts.

Overview of the proposed framework for community emotion aggregation and burst detection.
  1. Emotion Classification: A Multinomial Naive Bayes classifier, trained via distant supervision on emoticons, categorizes posts into four primary emotions: Happy, Anger, Sad, and Fear.
  2. Community Aggregation: To ensure robustness against low-authority "spammers," the framework employs the HITS algorithm to calculate an "emotional authority" score for each user within their social network. Aggregated community emotion is then calculated as a weighted sum of individual states, ensuring that influential community members have a greater impact on the detected signal.
  3. Burst Detection: The sequence of emotional states is modeled using a Dirichlet distribution to capture the probability distribution across the four emotion categories. A dynamic programming approach identifies transitions between "stable" and "burst" states, allowing the system to detect nuanced shifts, such as a simultaneous spike in both sadness and fear.

Experimental Results and Insights

The framework was evaluated using a massive dataset of 212 million Sina Weibo messages across four specialized communities: Legal Cases, Football, Economy, and Singers. Ground truth was established by cross-referencing detected bursts with major news archives.

Temporal distribution of emotions in March 2011, illustrating a clear 'Sad' emotion burst corresponding to a major real-world event.

The system achieved an F-measure of 0.683, significantly outperforming baseline methods like multi-stream burst detectors (0.632) and simple peak-finding (0.417). A key finding was the impact of authority weighting; incorporating HITS-based scores improved performance from 0.605 to 0.683, confirming that influential community members provide a more reliable signal for event detection. Qualitatively, the model successfully identified specific events like the "Qian Yunhui case" and Qatar's successful 2022 World Cup bid by tracking their unique emotional signatures.

Conclusion and Future Work

By shifting the focus from keyword frequency to community-weighted emotional distributions, this research provides a more nuanced and robust method for microblog event detection. While the current model assumes static authority, future work will explore dynamic authority modeling and finer-grained emotional lexicons to further enhance detection sensitivity for subtle sub-community events.

Summary updated

BibTeX
@inproceedings{ou2014exploiting,
  author = {Gaoyan Ou and Wei Chen and Tengjiao Wang and Zhongyu Wei and Binyang Li and Dongqing Yang and Kam-Fai Wong},
  title = {{Exploiting Community Emotion for Microblog Event Detection}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2014},
  doi = {10.3115/v1/d14-1123},
}
Figure