Detecting Hashtag Hijacking for Hashtag Activism

Pooneh Mousavi, Jessica OuyangACL 2021

TL;DR

This paper introduces a novel, weakly-supervised framework to detect "hashtag hijacking" specifically within social activism movements on Twitter, like #MeToo. Their system continuously updates itself to adapt to evolving hijacking tactics, outperforming existing methods by focusing on the dynamic nature of social movement discourse.

What Problem Does This Paper Address?

Social media has become a powerful tool for social movements, with hashtags like #MeToo and #BlackLivesMatter amplifying voices and raising awareness. However, these popular hashtags are also prime targets for "hijacking," where spammers or malicious actors co-opt the hashtag to promote unrelated content, spread misinformation, or even push hate speech. This not only dilutes the original message but can also harm marginalized groups. Existing methods for detecting hashtag hijacking often focus on general trending hashtags and struggle to adapt to new hijacking strategies or the nuanced evolution of social movement discourse. This paper aims to fill that gap by developing a system that can specifically identify hijacked tweets within the context of social activism and adapt over time.

Methods

The core of the proposed solution is a weakly-supervised, bootstrapping framework designed for continuous updating. It consists of two main components:

  1. Hijacked Tweet Classification Module: This module is an ensemble of five classifiers that work together to assign a "hijacked" score to each tweet. These classifiers include:
  2. Known Users Classifier: Maintains blacklists and whitelists of users based on their past posting behavior to identify known hijackers or trusted users.
  3. Tweet Text Classifier: Utilizes TF-IDF features from the tweet's text to identify patterns associated with hijacking.
  4. Social Classifier: Analyzes tweet interactions, such as retweets and likes, and user network features.
  5. User Profile Classifier: Examines user profile descriptions, hypothesizing that active participants in social movements will have more relevant profiles.
  6. Ensemble Voting: Combines the scores from the individual classifiers using strategies like Simple Average, Majority Vote, and a Stacking Meta-Learner. The Stacking Meta-Learner is highlighted for its ability to adapt to changing classifier performance.
  7. Batch Update Module: This crucial component allows the system to adapt over time. It periodically selects confidently labeled tweets (both hijacked and valid) from recent activity and adds them to the training data. This process then retrains the classifiers, enabling the system to learn new hijacking tactics and adapt to shifts in hashtag usage. The paper explores different update strategies, with "Update Equal" (preserving class balance) showing the most promise.

Datasets

  • #MeToo Tweets: The authors collected a new dataset of #MeToo tweets spanning from October 2017 to May 2020. This dataset was curated to include both potentially hijacked tweets (identified using spammy hashtags) and genuine tweets from the movement.
  • Seed Training Set: A seed set of 2770 #MeToo tweets was created for initial training and validation.
  • Expert-Labeled Data: To address issues with crowdsourced labels (low inter-annotator agreement), the authors used expert annotators to relabel a portion of the data, creating more reliable validation and test sets.
  • Live Tweets: For evaluating the batch update module, a large set of 122,792 live #MeToo tweets from February to May 2020 were collected.

Key Insights and Ideas

  • Adaptability is Key: The most significant insight is that hashtag hijacking strategies evolve. A static detection system will quickly become obsolete. The continuous updating mechanism is therefore critical for long-term effectiveness.
  • Social Movements are Dynamic: Unlike general trending hashtags, social activism hashtags have a fluid meaning and can evolve. The system's ability to adapt to these shifts (e.g., criticism of the movement being considered "related" rather than "hijacked") is a major advantage.
  • Ensemble Power: Combining multiple types of features (text, user, social) through an ensemble of classifiers, particularly with a meta-learner, significantly improves performance over individual classifiers.
  • Weak Supervision is Practical: For large-scale, evolving problems like this, relying solely on extensive manual annotation is infeasible. The weakly-supervised approach, leveraging initial noisy labels and bootstrapping, is a practical necessity.
  • Distinguishing Hijacking from Spam: The paper emphasizes that not all hijacked tweets are spam, and vice-versa. Their focus is on the intent of hijacking a specific hashtag's context, which is distinct from general spam detection.

Key Results

  • The proposed framework, particularly when using the Stacking Meta-Learner with the "Update Equal" strategy, significantly outperforms baselines, including Jain et al. (2015), on various metrics like ROC-AUC, Precision, Recall, and F-measure.
  • The seed-trained Stacking Meta-Learner achieved an F-measure of 0.814 on the Expert Test set.
  • When evaluating the batch update module on live data, the Stacking Meta-Learner with "Update Equal" achieved an F-measure of 0.722, demonstrating robust performance even with evolving data and potential noise.
  • The Tweet text classifier alone showed a significant drop in performance (about 0.16 F-measure) on live data without updates, highlighting the necessity of the update module and the benefit of other classifiers in the ensemble.

Limitations and Future Work

  • Class Imbalance and Error Propagation: The authors acknowledge that the inherent rarity of hijacked tweets in the wild can lead to class imbalance issues as the system updates. Error propagation, where incorrect labels from previous updates can negatively impact future performance, is also a concern.
  • User Blacklist Rigidity: The current blacklist mechanism might not allow users to be removed even if their posting behavior changes. The authors suggest adding criteria for users to be removed from the blacklist.
  • Future Work:
    • Investigating methods to address class imbalance and error propagation.
  • Exploring active learning with submodular data subset selection to efficiently obtain human-labeled samples for periodic retraining.
  • Incorporating additional features, such as analyzing linked web pages or embedded images, to further enhance classifier performance.
  • Developing other weakly-supervised approaches for Twitter data analysis.

Summary updated

BibTeX
@inproceedings{mousavi2021detecting,
  author = {Pooneh Mousavi and Jessica Ouyang},
  title = {{Detecting Hashtag Hijacking for Hashtag Activism}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure