It's not just sentiment classification. Many of these NLP methods overfit to the minutiae of the particular SOTA data set and data type. Language identificatio

It's not just sentiment classification. Many of these NLP methods overfit to the minutiae of the particular SOTA data set and data type. Language identification, NER, POS tagging, parsing, tokenization… you name it. Solved in theory, trouble in practice. Using the algos on a concrete project, with a concrete "upstream" goal, is the only way to assess the gap between SOTA and your needs. Don't blindly trust these 90+% scores.