Unsupervised learning of semantic orientation from a hundred-billion-word corpus

The evaluative character of a word is called its semantic orientation. A positive semantic orientation implies desirability (e.g., "honest", "intrepid") and a negative semantic orientation implies undesirability (e.g., "disturbing", "superfluous"). This paper introduces a simple algorithm for unsupervised learning of semantic orientation from extremely large corpora. The method involves issuing queries to a Web search engine and using pointwise mutual information to analyse the results. The algorithm is empirically evaluated using a training corpus of approximately one hundred billion words - the subset of the Web that is indexed by the chosen search engine. Tested with 3,596 words (1,614 positive and 1,982 negative), the algorithm attains an accuracy of 80 percent. The 3,596 test words include adjectives, adverbs, nouns, and verbs. The accuracy is comparable with the results achieved by Hatzivassiloglou and McKeown (1997), using a complex four-stage supervised learning algorithm that is restricted to determining the semantic orientation of adjectives.Le caract\ue8re d'\ue9valuation d'un mot est appel\ue9 son orientation s\ue9mantique. Une orientation s\ue9mantique positive suppose qu'un mot a un caract\ue8re d\ue9sirable (p. ex. "honn\ueate", "intr\ue9pide"), et une orientation s\ue9mantique n\ue9gative suppose qu'il a un caract\ue8re ind\ue9sirable (p. ex "d\ue9rangeant", "superflu"). Le pr\ue9sent document propose un algorithme simple permettant l'apprentissage non supervis\ue9 de l'orientation s\ue9mantique d'un corpus extr\ueamement volumineux. La m\ue9thode propos\ue9e consiste \ue0 interroger un moteur de recherche Web et \ue0 utiliser de l'information mutuelle ponctuelle pour analyser les r\ue9sultats. L'algorithme est \ue9valu\ue9 de fa\ue7on empirique au moyen d'un corpus d'apprentissage constitu\ue9 d'une centaine de milliards de mots environ, soit le sous ensemble du Web qui est index\ue9 par le moteur de recherche. L'algorithme, qui a \ue9t\ue9 test\ue9 avec 3 596 mots (1 614 mots positifs et 1 982 mots n\ue9gatifs), est pr\ue9cis \ue0 80 %. Les 3 596 mots test\ue9s comprennent des adjectifs, des adverbes, des noms et des verbes. L'exactitude des r\ue9sultats est comparable \ue0 celle des r\ue9sultats obtenus par Hatzivassiloglou et McKeown (1997) \ue0 l'aide d'un algorithme d'apprentissage non supervis\ue9 \ue0 quatre stades restreint \ue0 la d\ue9termination de l'orientation s\ue9mantique des adjectifs.Peer reviewed: NoNRC publication: Ye