Factual Knowledge

We propose DecoderLens, a method to interpret the iterative refinement of representations in encoder-decoder Transformer models.