Size matters: tight and loose context definitions in English word space models

摘要

Word Space Models use distributional similarity between two words as a measure of their semantic similarity or relatedness. This distributional similarity, however, is influenced by the type of context the models take intoaccount. Context definitions range on a continuum from tight to loose, depending on the size of the context window around the target or the order of the context words that are considered. This paper investigates whether two generalways of loosening the context definition — by extending the context size from one to ten words, and by taking into account second-order context words — produce equivalent results. In particular, we will evaluate the performanceof the models in terms of their ability (1) to discover semantic word classes and (2) to mirror human associations.