Research question
"How do these papers use the Transformer architecture, and what changes between the original model and BERT?"
The papers are related, but they solve different problems. A comparison should separate the shared architecture from the training objective and downstream use.
The Transformer establishes the architecture
Vaswani et al. introduce an encoder-decoder model based on attention mechanisms, removing recurrence and convolutions from the sequence-transduction architecture.
Attention Is All You Need, PDF pp. 1-2BERT turns the encoder into a pretrained representation
Devlin et al. pretrain deep bidirectional Transformer representations on unlabeled text, using left and right context across all layers.
BERT, PDF p. 1The downstream setup changes
The original Transformer is trained for sequence transduction tasks. BERT is designed for fine-tuning, often with one additional output layer for tasks such as question answering and language inference.
BERT, PDF pp. 1-2