Before it
n-gram models counted exact word sequences and assigned zero probability to anything unseen.
The problem
Sparsity: most word sequences never appear in any corpus, however large.
What it introduced
Cluster words into classes automatically, so statistics are shared between similar words.
Why it matters
The first practical answer to sparsity by grouping words that behave alike, an idea embeddings later generalised.
What modern models inherit
The instinct that similar words should share statistical strength.
What came next
Bengio's neural language model replaced discrete classes with continuous vectors.