A Look inside the Distributionally Similar Terms
Kow Kuroda, Jun’ichi Kazama, Kentaro Torisawa · 2010
We analyzed the details of aWeb-derived distributional data of Japanese nominal terms with two aims. One aim is to examine if distributionally similar terms can be in fact equated with “semanti-cally similar ” terms, and if so to what extent. The other is to investigate into what kind of semantic relations con-stitute (strongly) distributionally similar terms. Our results show that over 85% of the pairs of the terms derived from the highly similar terms turned out to be semantically similar in some way. The ratio of “classmate, ” synonymous, hypernym-hyponym, and meronymic re-lations are about 62%, 17%, 8 % and 1% of the classified data, respectively. 1