An Analysis of French-Language Tweets About COVID-19 Vaccines: Supervised Learning Approach

Area of Science:

• Natural Language Processing
• Machine Learning
• Social Media Analysis

Background:

• COVID-19 pandemic fueled societal disinformation, impacting vaccine hesitancy.
• Social media analysis is crucial for understanding disinformation's impact on public health.
• Machine learning and NLP are essential for analyzing large social media datasets.

Purpose of the Study:

• To evaluate the CamemBERT model's effectiveness in classifying French vaccination-related tweets.
• To assess the model's ability to handle ambiguous, sarcastic, or irrelevant tweet content.

Main Methods:

• Extracted 901,908 French vaccination tweets (July-August 2021).
• Labeled ~2000 tweets for pro/con vaccination arguments and content type (scientific, political, social).
• Fine-tuned and tested the CamemBERT model, assessing performance with F1-scores and confusion matrices.

Main Results:

• Achieved 70.6% accuracy for pro/con classification and 90% for content type classification.
• Tweets under 170 characters were 1.86 times more likely to be misclassified.
• Model accuracy is influenced by the classification task and tweet content heterogeneity.

Conclusions:

• CamemBERT's accuracy depends on the classification category and tweet topic.
• Political discourse in vaccine debates reduces classification accuracy for less distinct categories.
• Tweet length can be used to improve classification accuracy, enhancing analysis of social media data.