Sequence to Sequence Learning with Neural Networks

This paper introduces a general-purpose encoder-decoder framework using Long Short-Term Memory (LSTM) networks for sequence-to-sequence learning. It demonstrates that LSTMs can map input sequences to output sequences even when the lengths are different, and shows strong performance in machine translation tasks. The core idea involves using one LSTM to encode the input sequence into a fixed-size vector, and another LSTM to decode the target sequence from that vector. This model significantly improves translation quality by capturing long-range dependencies in the input.