In particular embodiments, annotating each keyword of a plurality of keywords with one or more labels of a plurality of labels, comprising: for each label, computing a score for the keyword document corresponding to the keyword and the label using an annotation model; and annotating the keyword with a specific label where the keyword document corresponding to the keyword and the specific label have the highest or the lowest score. Constructing a classifier based on a plurality of training keywords. For each keyword of the plurality of keywords, for each label annotating the keyword, calculating a second index-wise product between a word count vector of the keyword document corresponding to the keyword and a word count vector of the label document corresponding to the label; and predicting whether the label annotating the keyword is correct using the classifier with the second index-wise product as an input to the classifier.