New Experiments in Distributional Representations of Synonymy

摘要

Recent work on the problem of detectingsynonymy through corpus analysis hasused the Test of English as a Foreign Language(TOEFL) as a benchmark. However,this test involves as few as 80 questions,prompting questions regarding thestatistical significance of reported results.We overcome this limitation by generatinga TOEFL-like test using WordNet, containingthousands of questions and composedonly of words occurring with sufficientcorpus frequency to support sounddistributional comparisons. Experimentswith this test lead us to a similarity measurewhich significantly outperforms thebest proposed to date. Analysis suggeststhat a strength of this measure is its relativerobustness against polysemy.