I have to use a neural network to classify whether some reviews of hotels are deceptive or truthful. I also have to use pre-trained word embeddings to fed the neural networks. So I can use Word2vec to get the word vectors from a way larger dataset of hotels reviews. However, Word2vec gives the possibility to use continuous bag-of-words and continuous skip-gram models for this task. Which one would be generally better for this specific task?
1 Answer
I think this article gives general idea of pros / cons between CBOW and Skip Gram,
According to Mikolov:
Skip-gram: works well with small amount of the training data, represents well even rare words or phrases.
CBOW: several times faster to train than the skip-gram, slightly better accuracy for the frequent words.
So without looking at data, assuming that you have large amount of data, CBOW might be my bet. However, if you have labels for "deceptive" and "truthful", it will be best to train both models, and use whichever outperforms via cross-validation.