Getting into bed with embeddings? A comparison of collocations and word embeddings for corpus-assisted discourse analysis

Abstract

This paper discusses two approaches for identifying lexical patterns in discourse, namely the corpus linguistic method of collocation analysis and the natural language processing method of word embeddings. While both approaches can identify lexical patterns, they approach the task with different underlying frameworks, and the extent to which their results resemble one another has not been directly compared. This study uses two corpora, five collocation measures, and two word embedding algorithms to generate such comparisons. Results generally support the notion that many word pairs with similar embeddings are collocates, and that, to a lesser extent, many collocates have similar word embeddings. However, a major difference is that word pairs with similar embeddings do not need to co-occur often, or at all. Moreover, systematic differences in the kinds of words highlighted between the two word embedding algorithms were found and are discussed.