2018 · Representations in Context

Contextual Embeddings

The word stays the same. Its vector does not.

How it works

Word2Vec gives every word one permanent vector. That means bank beside a river and bank holding a loan are literally identical inside the model. ELMo changed the lookup into a computation: read the sentence in both directions, then build each word's representation from what surrounds it.

This laptop-sized demonstration keeps that conceptual move and strips away the large bidirectional network. It starts from real co-occurrence vectors learned from the sonnets, then blends the target with the current ±2-word context. The static vector is unchanged; the contextual result bends toward the words present now.

fair

thy fair face is bright

fair

a fair judgement and honest mind

The first twelve dimensions of two context-conditioned vectors. Different neighbours pull the same static word in different directions.

Try it — change the context

Where it falls short

The context rule is fixed. ELMo learns how syntax and meaning should change a vector; this demo averages neighbouring evidence.

The corpus is tiny. Rare meanings have little statistical support.

A window is not a sentence. Long-range dependencies still fall outside ±2 words.