2 ms·

Art of NLP: Classical, Modern and Contemporary Approaches to Text Classification

• PaulHoule 4y ago
  They divide the task into two parts: (1) generating a set of features, and (2) classifying based on the features. For (1) they find the word vectors and LSTM do not outperform bag-of-words which does not surprise me (bow works very well, new models seem to model the superficial aspects of text but don't capture semantics... they are a big pile of biases on top of biases), but they think the hyperparameter tuning to get there is easier with the newer models. For (2) they find extreme gradient boosting greatly outperforms the alternatives in terms of accuracy although performance in terms of speed is a lot worse. That was new to me. I would have liked to see some testing of dimensional reduction w/ bow. Also it irks me that it is widespread to compare a bunch of AUC and accuracy numbers but nobody puts error bars or statistical significance on them, but it's certainly true that if you repeated the experiment multiple times with different test/train splits you would get slightly different results.