Im making Sentiment Analysis from more than 8 languages using XLM-RoBERTa, the problem is, i have over 130.000 rows in my training dataset and lots of the 'Sentiment label (2=positive, 1=neutral, 0=negative)' are incorrect, can XLM-RoBERTa handle those noisy label? or should i just relabel it manually? I tried to use VADER to correct my labels automatially, but it doesnt seems helping, hence i need to know whether its better to just use the noisy label directly into the model, or should i correct the label manually Continue reading...
