The paper explores six key challenges for neural machine translation (NMT): domain mismatch, amount of training data, rare words, long sentences, word alignment, and beam search. It compares the performance of NMT and traditional statistical machine translation (SMT) systems on these challenges. Key findings include: 1. **Domain Mismatch**: NMT systems perform poorly out of domain, often sacrificing adequacy for fluency. 2. **Amount of Training Data**: NMT systems require more training data to achieve better performance compared to SMT systems, but they can still perform well with limited data. 3. **Rare Words**: NMT systems generally outperform SMT systems in translating very infrequent words, but both systems struggle with highly inflected categories. 4. **Long Sentences**: NMT systems are better at handling shorter sentences but perform poorly on very long sentences due to their tendency to produce shorter translations. 5. **Word Alignment**: The attention model in NMT does not fully function as a word alignment mechanism, leading to some divergence from traditional word alignment methods. 6. **Beam Search**: Increasing the beam size does not consistently improve translation quality and can even lead to worse translations beyond an optimal setting. The paper concludes that while NMT has shown significant progress, it still faces substantial challenges, particularly in out-of-domain and low-resource settings.The paper explores six key challenges for neural machine translation (NMT): domain mismatch, amount of training data, rare words, long sentences, word alignment, and beam search. It compares the performance of NMT and traditional statistical machine translation (SMT) systems on these challenges. Key findings include: 1. **Domain Mismatch**: NMT systems perform poorly out of domain, often sacrificing adequacy for fluency. 2. **Amount of Training Data**: NMT systems require more training data to achieve better performance compared to SMT systems, but they can still perform well with limited data. 3. **Rare Words**: NMT systems generally outperform SMT systems in translating very infrequent words, but both systems struggle with highly inflected categories. 4. **Long Sentences**: NMT systems are better at handling shorter sentences but perform poorly on very long sentences due to their tendency to produce shorter translations. 5. **Word Alignment**: The attention model in NMT does not fully function as a word alignment mechanism, leading to some divergence from traditional word alignment methods. 6. **Beam Search**: Increasing the beam size does not consistently improve translation quality and can even lead to worse translations beyond an optimal setting. The paper concludes that while NMT has shown significant progress, it still faces substantial challenges, particularly in out-of-domain and low-resource settings.