This paper investigates the properties of neural machine translation, a new approach to statistical machine translation based on neural networks. The authors focus on two models: an RNN Encoder-Decoder and a newly proposed gated recursive convolutional neural network (grConv). They evaluate these models on English-to-French translation tasks, analyzing their performance with respect to sentence length and the number of unknown words. The results show that neural machine translation performs well on short sentences without unknown words but degrades significantly as sentence length and the number of unknown words increase. The grConv model is found to learn the grammatical structure of sentences automatically, which is an interesting property for natural language processing applications. The paper concludes with suggestions for future research, including scaling up training, improving performance on long sentences, and exploring different neural architectures.This paper investigates the properties of neural machine translation, a new approach to statistical machine translation based on neural networks. The authors focus on two models: an RNN Encoder-Decoder and a newly proposed gated recursive convolutional neural network (grConv). They evaluate these models on English-to-French translation tasks, analyzing their performance with respect to sentence length and the number of unknown words. The results show that neural machine translation performs well on short sentences without unknown words but degrades significantly as sentence length and the number of unknown words increase. The grConv model is found to learn the grammatical structure of sentences automatically, which is an interesting property for natural language processing applications. The paper concludes with suggestions for future research, including scaling up training, improving performance on long sentences, and exploring different neural architectures.