Segment embedding

One of three embeddings summed to form each token's input representation. Tells the model whether a token belongs to sentence A or sentence B in a two-sentence input. Two learned vectors: E_A and E_B.