Abstract
In this paper, we deal with the task of sub-event detection in evolving events using posts collected from the Twitter stream. By representing a sequence of successive tweets in a short time interval as a weighted graph-of-words, we are able to identify the key moments (sub-events) that compose an event using the concept of graph degeneracy. We then select a tweet to best describe each sub-event using a simple yet effective heuristic. We evaluated our approach using human-generated summaries containing the actual important sub-events within each event and compare it to two baseline approaches using several performance metrics such as DET curves and precision/recall performance. Extensive experiments on recent sporting event streams indicate that our approach outperforms the dominant sub-event detection methods and constructs a human-readable event summary by aggregating the most representative tweets of each sub-event.
Introduction
Twitter serves as a vital real-time information source, yet its inherent noise makes extracting specific "sub-events"—pivotal moments within a larger ongoing event—a significant challenge. While detecting a major event like an earthquake is relatively straightforward via volume spikes, identifying a yellow card or a specific goal in a soccer match requires more nuanced analysis. The 140-character limit and irregular language further complicate traditional natural language processing (NLP) approaches. This paper proposes moving beyond simple frequency counts to analyze the structural evolution of conversation using graph degeneracy.
Methodology
The proposed pipeline consists of three main stages: feature extraction, sub-event detection, and tweet selection.
Feature Extraction and Graph-of-Words
Instead of treating tweets as bags of words, they are represented as a weighted graph-of-words. Nodes represent unique terms, and edges represent co-occurrence within a tweet. Edge weights are normalized based on the number of terms in the tweet to ensure that long and short tweets are treated fairly.
Graph Degeneracy (k-cores)
The system applies k-core decomposition to identify dense, cohesive clusters of words. A k-core is a maximal subgraph where every node has at least degree k. During a sub-event, specific keywords (e.g., "goal", "Neymar") suddenly co-occur frequently, forming a highly dense core that stands out from background chatter.
Detection and Selection
Sub-events are triggered when the sum of core numbers for the top-10 most central words significantly exceeds a moving average threshold ($\theta$). Once detected, the system selects the single tweet from that window with the highest total "core weight" to serve as a summary of the moment.
Experimental Evaluation
The approach was validated using a large-scale dataset from the 2014 FIFA World Cup, comprising approximately 7.5 million tweets across 11 matches, including the final. Ground truth sub-events were established using professional live match reports from ESPN FC. The system was also tested against a benchmark 2010 dataset to ensure a fair comparison with existing literature.
Results and Discussion
The proposed "Weight-Core" method demonstrated superior performance, achieving a Micro-average F1-score of 0.68, compared to 0.54 for the standard frequency-based baseline. Notably, the system detected 30 out of 32 goals in the 2014 dataset and all match-end moments. On the 2010 dataset, it achieved an F1-score of 0.93, significantly outperforming previous state-of-the-art methods (0.73).
Key insights include:
- Cohesion over Volume: The density of vocabulary is a more reliable signal than raw tweet volume for identifying sub-events.
- Unsupervised Nature: The method requires no training data or domain-specific knowledge, making it robust for various event types.
- Noise Filtering: Graph degeneracy naturally filters out random chatter, as non-informative words do not form dense clusters.
Conclusion
By focusing on vocabulary cohesion rather than raw volume, this unsupervised method effectively filters noise and identifies meaningful sub-events in real-time. While the system occasionally missed events due to simultaneous noise (e.g., a red card occurring during goals in another match), future work will explore adaptive time windows to better capture events with varying temporal dynamics.
Summary updated