Neural Machine Translation by Jointly Learning to…

This paper asks how a translation model can look back at the right source words instead of compressing an entire sentence into one fixed vector. Read it to see how a bidirectional encoder leaves one annotation at every source position, how a jointly trained alignment scorer builds a fresh weighted context for each target word, and why strong WMT14 results still need careful treatment of unknown words, sequence length, exhaustive scoring, and causal claims.

Reading focus: Why one fixed context vector can become a bottleneck, and how per-position source annotations keep the sentence addressable. How alignment scores, softmax weights, and a weighted sum produce a different context vector for every target word. What the WMT14 BLEU and long-sentence results establish, and where vocabulary, compute, and architectural confounders limit the conclusion.

ICLR 2015. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 30 min read, easy difficulty.