By the way, I am seeing a ~20% accuracy gain over TextBlob using this pre-trained Transformers approach, on the Kaggle 10k IMdB movie reviews dataset.