Rumour identification on Twitter as a function of novel textual and language-context features

Area of Science:

• Computer Science
• Information Science
• Social Media Analysis

Background:

• Social microblogs facilitate rapid information dissemination but also serve as conduits for rumor propagation.
• Existing rumor detection methods often operate at the topic level, lacking tweet-specific granularity.
• Prior research has limitations in utilizing discrete emotions and effective part-of-speech features within content-based approaches.

Purpose of the Study:

• To develop a robust framework for rumor detection at the individual tweet/post level.
• To integrate both context-based and content-based features for enhanced rumor identification.
• To address the limitations of previous studies in feature representation for rumor detection.

Main Methods:

• A novel framework combining word2vec embeddings and Bidirectional Encoder Representations from Transformers (BERT) for context-based features.
• Incorporation of discrete emotions, linguistic, and metadata characteristics for content-based features.
• Utilized four real-life Twitter microblog datasets for comprehensive testing and validation.

Main Results:

• The proposed framework achieved high rumor detection rates of 97%, 86%, 85%, and 80% across four diverse datasets.
• The model significantly outperformed three state-of-the-art baseline approaches in rumor detection accuracy.
• Bidirectional Encoder Representations from Transformers (BERT) demonstrated superior performance among context-based methods, while linguistic features excelled in content-based approaches.

Conclusions:

• The developed framework offers a significant advancement in tweet-level rumor detection.
• The integration of diverse features, including context and content, is crucial for effective misinformation identification.
• Further improvements in detection model performance were observed through the application of a two-step feature selection process.