A closer look at the word analogy problem

Siddharth Krishna Kumar

Abstract: Although word analogy problems have become a standard tool for evaluating word vectors, little is known about why word vectors are so good at solving these problems. In this paper, I attempt to further our understanding of the subject, by developing a simple, but highly accurate generative approach to solve the word analogy problem for the case when all terms involved in the problem are nouns. My results demonstrate the ambiguities associated with learning the relationship between a word pair, and the role of the training dataset in determining the relationship which gets most highlighted. Furthermore, my results show that the ability of a model to accurately solve the word analogy problem may not be indicative of a model’s ability to learn the relationship between a word pair the way a human does.

TL;DR: Simple generative approach to solve the word analogy problem which yields insights into word relationships, and the problems with estimating them

Keywords: word2vec, glove, word analogy, word relationships, word vectors