This paper proposes a novel neural machine translation approach that jointly learns to align and translate. Unlike traditional encoder-decoder models that compress a source sentence into a fixed-length vector, the proposed method allows the model to automatically search for relevant parts of the source sentence without explicitly forming them. This approach enables the model to handle long sentences more effectively by focusing on the most relevant information for each target word. The model achieves translation performance comparable to existing state-of-the-art phrase-based systems on English-to-French translation. Qualitative analysis shows that the (soft-)alignments found by the model align well with human intuition. The proposed approach uses a bidirectional RNN as an encoder to generate annotations of the source sentence, which are then used by the decoder to compute context vectors for each target word. The decoder uses these context vectors to predict the next target word, allowing it to focus on the most relevant parts of the source sentence. This mechanism enables the model to handle long sentences more effectively by avoiding the need to encode the entire sentence into a fixed-length vector. The model was evaluated on the task of English-to-French translation using the WMT '14 corpus. The results show that the proposed approach outperforms the conventional encoder-decoder model in terms of translation quality, especially for longer sentences. The model is also more robust to the length of the source sentence. Qualitative analysis of the model's alignments shows that the (soft-)alignments are linguistically plausible and align well with human intuition. The proposed approach represents a significant advancement in neural machine translation, as it allows the model to handle long sentences more effectively by focusing on the most relevant parts of the source sentence. This approach is a promising step toward better machine translation and a better understanding of natural languages.This paper proposes a novel neural machine translation approach that jointly learns to align and translate. Unlike traditional encoder-decoder models that compress a source sentence into a fixed-length vector, the proposed method allows the model to automatically search for relevant parts of the source sentence without explicitly forming them. This approach enables the model to handle long sentences more effectively by focusing on the most relevant information for each target word. The model achieves translation performance comparable to existing state-of-the-art phrase-based systems on English-to-French translation. Qualitative analysis shows that the (soft-)alignments found by the model align well with human intuition. The proposed approach uses a bidirectional RNN as an encoder to generate annotations of the source sentence, which are then used by the decoder to compute context vectors for each target word. The decoder uses these context vectors to predict the next target word, allowing it to focus on the most relevant parts of the source sentence. This mechanism enables the model to handle long sentences more effectively by avoiding the need to encode the entire sentence into a fixed-length vector. The model was evaluated on the task of English-to-French translation using the WMT '14 corpus. The results show that the proposed approach outperforms the conventional encoder-decoder model in terms of translation quality, especially for longer sentences. The model is also more robust to the length of the source sentence. Qualitative analysis of the model's alignments shows that the (soft-)alignments are linguistically plausible and align well with human intuition. The proposed approach represents a significant advancement in neural machine translation, as it allows the model to handle long sentences more effectively by focusing on the most relevant parts of the source sentence. This approach is a promising step toward better machine translation and a better understanding of natural languages.