Recently Facebook, after a three year delay, has finally banned QAnon, the far-right conspiracy theory from its’ platform. Following their lead, Twitter and Youtube have followed suit. Grown from the roots of white supremacy, what was once a theory has has morphed into an aggressive movement seeking to wage information warfare as a means of influencing the upcoming American presidential election in favour of Donald Trump. The FBI has identified QAnon as a domestic terrorism threat, and earlier this year Twitter acted to ban thousands of QAnon-affiliated accounts. However, many pro-QAnon accounts continue to tweet, and more insidiously Trump has a proven history of amplifying these voices by irresponsibly retweeting their malicious content. He also refuses to publicly condemn the theory, side-stepping with comments about pedophilia, further spreading toxic uncertainty.

I thought it would be interesting to investigate what these accounts were tweeting. Specifically, I was curious about the most popular topics discussed by the far-right and whether I could leverage recent developments in machine learning to analyze tweets from currently active, known QAnon accounts on Twitter. I decided to try zero-shot learning to classify tweets into topic categories of my choice: politics, public health, climate change, religion, controversy and fake news. For this experiment I used the Twitter API to gather tweets, and Hugging Face’s out-of-the-box implementation of zero-shot learning for text classification (released August, 2020). As per an unsupervised learning approach, unlabeled tweets were classified into socially relevant topics; without any task specific training or annotated tweets for fine-tuning. To make my zero-shot classifier accessible, I built a web app with the Python framework, Streamlit, for an interactive interface where I could drag and drop a large batch of unprocessed tweets stored as a csv or json file.

#nlp #misinformation #social-media #machine-learning