A method including: generating, from a text corpus, a lexicon of unigrams and bigrams comprising an embedding for each of said unigrams and bigrams; training a machine learning classifier on a training set comprising a subset of said lexicon, wherein each of said unigrams and bigrams in said subset has a sentiment label; applying said machine learning classifier to said lexicon, to (i) predict a sentiment of each of said unigrams and bigrams, and (ii) update said lexicon with the predicted sentiments; and performing statistical analysis on said updated lexicon, to extract one or more sentiment composition lexicons, wherein each of said one or more sentiment composition lexicons is associated with a sentiment composition class.