Twitter Summarization Based on Social Network and Sparse Reconstruction

Ruifang He, Xingyi DuanAAAI 2018 · DOI

Abstract

Summarizing microblogging content like Twitter presents unique challenges due to the brevity, noise, and high redundancy of the data. Traditional text-centric methods often fail to capture the underlying semantic structure of social conversations. This paper introduces SNSR (Social Network and Sparse Reconstruction), a framework that integrates social network theories with sparse reconstruction to produce high-quality summaries. By modeling "expression consistency" and "expression contagion," SNSR leverages user relationships to identify salient information. Experimental results on a real-world Twitter dataset demonstrate that SNSR significantly outperforms traditional baselines, with social regularization providing a substantial boost to summary quality.

Introduction

Twitter has become a primary source for real-time information, yet its "firehose" nature makes it difficult for users to digest large volumes of tweets on specific topics. Summarization is essential, but standard extractive techniques (e.g., LexRank) treat tweets as independent text units, ignoring the rich social context. Twitter data is characterized by two main hurdles: the sparsity of text in individual tweets and the high redundancy caused by "echo chambers." To address these, the SNSR model views tweets not just as strings, but as nodes in a social graph.

The SNSR Framework

The SNSR model combines three distinct components into a unified optimization objective:

1. Sparse Reconstruction (SR)

The core of the model is finding a small subset of tweets that can mathematically "reconstruct" the entire corpus. This is formulated as a Group Lasso problem, which selects a sparse set of representative tweets that minimize the reconstruction error of the original dataset.

2. Social Regularization (SN)

To enhance the reconstruction, the framework incorporates two key social theories:

  • Expression Consistency: Users tend to maintain consistent sentiments and preferences within a short timeframe.
  • Expression Contagion: Social ties influence expression; friends are more likely to share similar viewpoints and vocabulary. By regularizing the model with these social links, tweets from the same user or connected friends are treated as semantically related, helping the model identify core themes that transcend individual word choices.

3. Diversity Regularization

To ensure the summary covers a broad range of perspectives, a diversity penalty is applied. Using a cosine similarity threshold, the model penalizes the selection of tweets that are too similar to those already included in the summary.

Experimental Evaluation

The authors evaluated SNSR on a custom CTS (Twitter Summarization) Dataset consisting of 12 major events from 2012, such as the Joplin tornado and the death of Osama bin Laden. Gold-standard summaries were created by 24 expert volunteers to provide a ground truth for ROUGE evaluation.

Results and Discussion

SNSR demonstrated superior performance across all standard metrics (ROUGE-1, ROUGE-2, and ROUGE-SU4) compared to traditional methods like LexRank, LSA, and other sparse-reconstruction variants like DSDR.

Table 2: Performance comparison of SNSR against various baseline summarization methods.

The inclusion of social information proved to be the most critical innovation. An ablation study revealed that social regularization specifically improved ROUGE-2 scores by 46.01%. This suggests that social context is highly effective at identifying meaningful bigrams and phrases that text-only models might miss.

Impact of different regularization components on ROUGE scores.

Conclusion

The SNSR framework successfully bridges the gap between social network analysis and sparse optimization. By treating the social graph as a source of semantic evidence, the model overcomes the inherent noise and brevity of Twitter data. Future work may involve exploring dynamic social graphs and automated hyperparameter tuning to further refine real-time summarization capabilities.

Summary updated

BibTeX
@inproceedings{he2018twitter,
  author = {Ruifang He and Xingyi Duan},
  title = {{Twitter Summarization Based on Social Network and Sparse Reconstruction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.12058},
}
Figure