Improving classification of tweets using word-word co-occurrence information from a large external corpus.