Word embeddings have gained a lot of traction in recent years since their multidimensional representation of words exhibits useful semantic properties while the process of obtaining them requires little to no manual interaction. At the same time, linguistic theories of lexical semantics have been under active research for many decades and have rich and well-established foundations. Have such theories now become obsolete with the advent of sophisticated statistical methods based on machine learning? This paper investigates the relationship between a number of widely used variants of word embeddings and one specific theory of lexical semantics, viz. frame semantics. Based on the lexical units available in the FRAMENET corpus, we study some commonalities and differences in the way these are represented in each formalism. We comment on possible future directions that allow both worlds to learn from each other and thus may help improve practical aspects of word embeddings as well as frame semantics.