Timeline Summarization from Social Media with Life Cycle Models

Yi Chang, Jiliang Tang, Dawei Yin, Makoto Yamada, Yan Liu

Abstract

Generating chronological summaries from social media data presents unique challenges due to the informal, noisy, and high-volume nature of platforms like Twitter. This paper introduces Timeline-Sumy, a framework that models the "life cycle" of events—characterized by sudden spikes and long tails of interest—to improve episode discovery and content selection. By integrating Bayesian nonparametric modeling with a Learning-to-Rank approach, Timeline-Sumy significantly outperforms traditional text-based summarization methods.

Introduction

Summarizing social media is a "needle in a haystack" problem. Unlike structured news archives, social media posts are brief and informal. Traditional tools often fail because they rely solely on textual content, ignoring the temporal dynamics and social signals inherent in these platforms. The authors identify two primary hurdles: Episode Discovery (identifying significant events amidst constant chatter) and Content Quality (selecting the most representative post for an episode).

Methodology

Timeline-Sumy addresses these challenges through a two-phase process:

1. Episode Detection

The framework employs a Bayesian nonparametric model to group posts into episodes.

  • Temporal Modeling: A Gamma distribution captures the typical social media event life cycle: a sharp rise in interest followed by a lingering "heavy tail."
  • Content Modeling: Text and hashtags are modeled separately, recognizing that hashtags serve as sparse but highly informative metadata.
  • Automatic Scaling: Utilizing a Chinese Restaurant Process (CRP), the model dynamically determines the number of episodes without human intervention.

2. Summary Ranking

Once episodes are identified, a Learning-to-Rank (LTR) approach using Gradient Boosted Decision Trees (GBDT) selects the best representative post. Features include:

  • Temporal: Proximity to the episode's peak interest.
  • Textual: TF-IDF similarity, post length, and language quality.
  • Social Popularity: Engagement metrics like retweets, likes, and follower counts.

To improve efficiency, a "Fast Burn-in" strategy using group lasso optimization reduces the iterations required for model convergence from over 500 to just 20.

Evaluation and Datasets

The framework was evaluated on two distinct datasets:

  1. Labeled Datasets: Approximately 685k posts regarding four tennis stars (Andy Murray, David Ferrer, Maria Sharapova, and Roger Federer) during the 2012 Wimbledon and London Olympics.
  2. Unlabeled Case Study: Over 600k posts about Jennifer Lopez from 2014 to demonstrate real-world utility in summarizing diverse events like concert performances and social media gossip.
Case study of the Jennifer Lopez dataset, showing detected episodes and representative posts.

Results and Discussion

Timeline-Sumy consistently outperformed baseline methods such as LexRank and ETS. As shown in the performance table below, the model achieved superior ROUGE scores across all tested datasets. For instance, in the Roger Federer dataset, it reached a ROUGE-L score of 0.53180, a substantial improvement over the next best baseline.

Comparison of ROUGE-2 and ROUGE-L scores across different datasets and methods.

Key insights from the study include:

  • Time as a Primary Signal: Modeling the shape of time-series data is more effective than treating timestamps as simple metadata.
  • Hashtag Separation: Treating hashtags as distinct from regular text preserves their high-signal value.
  • Popularity as Quality: Social engagement metrics are vital for identifying "representative" content in a noisy environment.

Conclusion

Timeline-Sumy demonstrates that explicitly modeling the temporal life cycle of social media events leads to more accurate and representative summaries. While currently focused on single entities, future work will explore multi-entity summarization to capture complex, intertwined events. Additionally, addressing challenges like bot-generated spam and misinformation remains a priority for future iterations of the framework.

Summary updated

Abstract

The popularity of social media shatters the barrier for online users to create and share information at any place at any time. As a consequence, it has become increasing difficult to locate relevance information about an entity. Timeline has been proven to provide an effective and efficient access to understand an entity by displaying a list of episodes about the entity in chronological order. However, summarizing the timeline about an entity with social media data faces new challenges. First, key timeline episodes about the entity are typically unavailable in existing social media services. Second, the short, noisy and informal nature of social media posts determines that only content-based summarization could be insufficient. In this paper, we investigate the problem of timeline summarization and propose a novel framework Timeline-Sumy, which consists of episode detecting and summary ranking. In episode detecting, we explicitly model temporal information with life cycle models to detect timeline episodes since episodes usually exhibit sudden-rise-and-heavy-tail patterns on time-series. In summary ranking, we rank social media posts in each episode via a learning-to-rank approach. The experimental results on social media datasets demonstrate the effectiveness of the proposed framework.
BibTeX
@inproceedings{chang2016timeline,
  author = {Yi Chang and Jiliang Tang and Dawei Yin and Makoto Yamada and Yan Liu},
  title = {{Timeline Summarization from Social Media with Life Cycle Models}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure