Questions for Self-monitoring: Encoder-Decoder Architecture and Attention

• Why can't machine translation be treated as sequence or segment labeling task?
• What sets a seq2seq model apart from the other architectures discussed so far?
• How is a seq2seq model trained?
• How do seq2seq models perform inference?

• What is the most serious drawback of the encoder-decoder architecture?
• How can this drawback be overcome?
• Why cannot all the information from the encoder be passed on to the decoder?
• What's the underlying idea of attention?
• How does dot-product attention work?