This paper analyzes the properties of neural machine translation (NMT) using two models: RNN Encoder–Decoder and a newly proposed gated recursive convolutional neural network (grConv). NMT is a new approach to statistical machine translation based purely on neural networks. It consists of an encoder and a decoder. The encoder extracts a fixed-length representation from a variable-length input sentence, and the decoder generates a correct translation from this representation. The paper shows that NMT performs relatively well on short sentences without unknown words but degrades rapidly as sentence length and the number of unknown words increase. Additionally, the grConv model is found to automatically learn a grammatical structure of a sentence. The paper discusses the properties of NMT models, focusing on their performance on sentences of varying lengths and with different numbers of unknown words. It is found that the performance of NMT models degrades significantly with longer sentences and a higher number of unknown words. However, both models are able to generate correct translations most of the time. The grConv model is able to learn a syntactic structure of the source language without supervision. The paper also presents an experiment on English-to-French translation using the two models. The results show that the grConv model performs better than the RNNenc model in some cases, particularly for shorter sentences. However, the RNNenc model still outperforms the grConv model in terms of BLEU scores for longer sentences. The analysis suggests that the current NMT approach has weaknesses in handling long sentences, likely due to the fixed-length vector representation not being able to encode complex structures and meanings. The paper concludes that further research is needed to improve NMT systems, particularly in handling long sentences and expanding vocabulary sizes. It also highlights the potential of the grConv model in learning grammatical structures automatically, making it suitable for natural language processing applications beyond machine translation.This paper analyzes the properties of neural machine translation (NMT) using two models: RNN Encoder–Decoder and a newly proposed gated recursive convolutional neural network (grConv). NMT is a new approach to statistical machine translation based purely on neural networks. It consists of an encoder and a decoder. The encoder extracts a fixed-length representation from a variable-length input sentence, and the decoder generates a correct translation from this representation. The paper shows that NMT performs relatively well on short sentences without unknown words but degrades rapidly as sentence length and the number of unknown words increase. Additionally, the grConv model is found to automatically learn a grammatical structure of a sentence. The paper discusses the properties of NMT models, focusing on their performance on sentences of varying lengths and with different numbers of unknown words. It is found that the performance of NMT models degrades significantly with longer sentences and a higher number of unknown words. However, both models are able to generate correct translations most of the time. The grConv model is able to learn a syntactic structure of the source language without supervision. The paper also presents an experiment on English-to-French translation using the two models. The results show that the grConv model performs better than the RNNenc model in some cases, particularly for shorter sentences. However, the RNNenc model still outperforms the grConv model in terms of BLEU scores for longer sentences. The analysis suggests that the current NMT approach has weaknesses in handling long sentences, likely due to the fixed-length vector representation not being able to encode complex structures and meanings. The paper concludes that further research is needed to improve NMT systems, particularly in handling long sentences and expanding vocabulary sizes. It also highlights the potential of the grConv model in learning grammatical structures automatically, making it suitable for natural language processing applications beyond machine translation.