Paraphrases do not explain word analogies
This work addresses a theoretical gap in understanding word embeddings for NLP researchers, but it is incremental as it refutes a specific claim without proposing a new solution.
The paper challenges a prior theoretical explanation that word analogies in embeddings arise from paraphrase relations, showing that the required paraphrases do not hold empirically.
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.