Sentiment analysis of tweets from US politicians with FastAI (natural language processing).
• Followed the ULMFiT methodology: Transfer learning with an AWD LSTM language model from Wikinet-103 dataset to the Sentiment-140 dataset. A language model predicts the next word/letter/symbol given a sequence.
• Trained the Sentiment-140 language model as a classifier, to predict sentiment of tweets from almost 2000 United States politicians.
• Classified the tweets by sentiment and analysed the data, comparing political parties, genders, birth countries and day by day averages. Found a significant decrease in average sentiment on the day of the Capitol attack (6th Jan 2021).