Language, like other natural sequences, exhibits statistical dependencies at a wide range of scales (Lin & Tegmark, 2016). However, many statistical learning models applied to language impose a sampling scale while extracting statistical structure. For instance, Word2Vec creates vector embeddings by sampling context in a window around each word, the size of which defines a strong scale; relationships over much larger temporal scales would be invisible to the algorithm. This paper examines the family of Word2Vec embeddings generated while systematically manipulating the size of the context window. The primary result is that different linguistic relationships are preferentially encoded at different scales. Different scales emphasize different syntactic and semantic relations between words, as assessed both by analogical reasoning tasks in the Google Analogies test set and human similarity rating datasets WordSim-353 and SimLex-999. Moreover, the neighborhoods of a given word in the embeddings change considerably depending on the scale. These results suggest that sampling at any individual scale can only identify a subset of the meaningful relationships a word might have, and point toward the importance of developing scale-free models of semantic meaning.