Something I don't understand:
I thought that when using LSTM for classification, the input vector for the sentence was a single vector of length 300, recurrently updated with each word in the text. And for the averaging method, it was a vector of length 300, the average of each of the vectors of length 300 for the words in the sentence. At what stage in the standard process would a matrix of size V*300 be used to represent a sentence or document?