2Note that the unsupervised Word2vec embeddings could
be used as input to a supervised classifier, which we did not
do. For a supervised embedding baseline we instead use WSABIE.
WSABIE trains word embeddings U and hashtag embeddings
V in a supervised fashion, whereasWord2vec trains
them both unsupervised. Adding supervision to Word2vec
would effectively do something in-between: U would still be
unsupervised, but V would then be supervised.