Paraphrases do not explain word analogies

TL;DR

This paper challenges a recent theoretical explanation for why word embeddings encode linguistic regularities, demonstrating that the paraphrase relations assumed in the explanation do not hold empirically.

Contribution

The paper refutes a recent theory claiming paraphrase relations explain word analogy regularities in embeddings, showing empirical evidence against this explanation.

Findings

Paraphrase relations do not hold empirically in the context of word analogies.

Theoretical explanation by Allen and Hospedales is not supported by empirical data.

Word embeddings encode regularities through mechanisms other than paraphrase relations.

Abstract

Many types of distributional word embeddings (weakly) encode linguistic regularities as directions (the difference between "jump" and "jumped" will be in a similar direction to that of "walk" and "walked," and so on). Several attempts have been made to explain this fact. We respond to Allen and Hospedales' recent (ICML, 2019) theoretical explanation, which claims that word2vec and GloVe will encode linguistic regularities whenever a specific relation of paraphrase holds between the four words involved in the regularity. We demonstrate that the explanation does not go through: the paraphrase relations needed under this explanation do not hold empirically.