when it comes to sentiment analysis, are there any algorithm that does not require manual labeling of training sentences "this one is negative", "this one is positive", ie unsupervised learning? From my understanding you can use BERT for sentiment analysis without labeling, just because it was trained on tons and tons of data, but original BERT training data had to be labelled, right? So if you try to predict some sentence, BERT just uses existing sentiment dictionary? Can someone explain this to me in more detail? Pic unrelated