Comparing Word Relatedness Measures Based on Google $n$-grams

Estimating word relatedness is essential in natural language processing (NLP), and in many other related areas. Corpus-based word relatedness has its advantages over knowledge-based supervised measures. There are many corpus-based measures in the literature that can not be compared to each other as they use a different corpus. The purpose of this paper is to show how to evaluate different corpus-based measures of word relatedness by calculating them over a common corpus (i.e., the Google n-grams) and then assessing their performance with respect to gold standard relatedness datasets. We evaluate six of these measures as a starting point, all of which are re-implemented using the Google n-gram corpus as their only resource, by comparing their performance in five different data sets. We also show how a word relatedness measure based on a web search engine can be implemented using the Google n-gram corpus.