COVID-19 and Misinformation: A Large-Scale Lexical Analysis on Twitter

Dimosthenis Antypas, Jose Camacho-Collados, Alun Preece, David RogersACL 2021

TL;DR

This paper analyzes the language used in COVID-19 misinformation tweets on Twitter, finding that misinformation tends to use a more limited and negative vocabulary than general tweets. They demonstrate that simple lexical features are surprisingly effective at identifying misinformation, achieving over 80% accuracy with a linear model.

What Problem Does This Paper Address?

The COVID-19 pandemic saw an unprecedented surge in misinformation on social media, posing a significant threat to public health. While the existence of misinformation is well-known, understanding the specific linguistic patterns that characterize it, especially in the context of a global health crisis, is crucial for developing effective detection and mitigation strategies. This paper aims to bridge this gap by performing a large-scale lexical analysis of COVID-19 misinformation on Twitter, seeking to identify distinct vocabulary patterns that differentiate it from legitimate discourse.

Methods

The researchers compiled two distinct Twitter corpora: one containing tweets identified as misinformation related to COVID-19, and a "generic" corpus of randomly selected COVID-19 tweets. Their analysis involved:

  • Lexical Feature Extraction: They extracted various linguistic features from tweets, including token counts, emoji usage, hashtags, user mentions, uppercase letters, punctuation, and exclamation marks.
  • Lexical Diversity Measurement: To assess vocabulary richness, they employed the Type-Token Ratio (TTR) and the Measure of Textual Lexical Diversity (MTLD).
  • Lexical Specificity Analysis: Using a hypergeometric distribution, they identified words that are statistically representative of either the misinformation or generic corpus, helping to pinpoint distinct vocabulary.
  • Machine Learning Classification: They built and evaluated several machine learning models (Naive Bayes, SVM, CNN, and BERT) using different combinations of lexical, semantic (word embeddings), and extra-linguistic features to classify tweets as either misinformation or generic.

Datasets

The study utilized two large Twitter corpora collected from January to April 2020, focusing on English-language tweets related to COVID-19.

  • Misinformation-Related Corpus: This corpus was extracted from a longitudinal study of misinformation call-outs, specifically targeting tweets that users tagged as "fake news," "disinformation," or "lies."
  • Generic Corpus: This corpus consisted of randomly selected COVID-19 related tweets, collected using the Tweepy library and a list of keywords.

Both corpora were balanced across months, with approximately 441,353 tweets in each. The authors note that while the generic set aims to be representative of general discourse, a small number of misinformation tweets might still be present.

Key Insights and Ideas

  • Misinformation Speaks a Different Language: The most striking finding is that misinformation tweets exhibit a significantly more limited and negative vocabulary compared to general COVID-19 tweets. This suggests that misinformation campaigns might rely on a narrower set of emotionally charged or fear-inducing terms.
  • Simplicity Can Be Powerful: Surprisingly, a simple linear model using only lexical features achieved an accuracy of over 80% in identifying misinformation. This highlights the strong signal embedded in the word choices themselves, even without considering complex linguistic structures or context.
  • User Mentions as a Signal: Misinformation tweets showed a disproportionately higher number of user mentions (@) compared to generic tweets. This could indicate a strategy of direct engagement or amplification within specific online communities.
  • Exclamation Marks for Emphasis: Misinformation tweets used exclamation marks more frequently, suggesting an attempt to convey urgency or strong emotion, a common tactic in persuasive or manipulative communication.
  • Subtopic Specificity: Analysis of specific subtopics like "Covid/Weapon" and "5G" conspiracy theories revealed distinct lexical patterns, with misinformation tweets in these areas often employing more intimidating or conspiratorial language.

Key Results

  • Lexical Diversity: The misinformation corpus demonstrated lower lexical diversity (MTLD score of 268.83) compared to the generic corpus (MTLD score of 593.74).
  • Classification Accuracy: A simple SVM model using TF-IDF and word embeddings achieved an accuracy of 0.82. The more advanced BERT model achieved the highest overall accuracy of 0.91.
  • Feature Effectiveness: The study confirmed that lexical features alone are highly effective, with linear models achieving over 80% accuracy.
  • Monthly Trends: Analysis across months showed that while BERT consistently outperformed SVM, both models maintained relatively stable performance, suggesting robustness in detecting misinformation over time.

Limitations and Future Work

The authors acknowledge several limitations:

  • Corpus Collection Bias: The misinformation corpus was collected based on users "calling out" misinformation. This "call out" principle might not capture all forms of misinformation and could introduce a bias.
  • "Generic" Corpus Imperfection: While efforts were made to create a clean generic corpus, it's acknowledged that a small number of misinformation tweets might still be present.
  • Model Robustness: While the models showed robustness across different months, the authors suggest that further evaluation on manually labeled datasets, rather than those collected via the "call out" principle, would be beneficial.
  • Future Directions: The paper suggests exploring other datasets and evaluating models on manually labeled data to confirm findings. It also implies that further investigation into the specific linguistic properties that contribute to misinformation's distinctiveness would be valuable.

Summary updated

BibTeX
@inproceedings{antypas2021covid,
  author = {Dimosthenis Antypas and Jose Camacho-Collados and Alun Preece and David Rogers},
  title = {{COVID-19 and Misinformation: A Large-Scale Lexical Analysis on Twitter}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure