文章结合机器翻译发展历程，着重关注了语境上下文信息在翻译过程中由于依赖关系而丢失这一问题，首先，探究了LSTM神经网络模型及其相关变种模型，建立了机器翻译编码器-解码器框架；接着，在Multi30K数据集上使用Pytorch深度学习框架实现了德文到英文的机器翻译功能；然后，采用BLEU评分作为模型评价指标，对比了LSTM与GRU等相关模型的实验结果。实验结果表明，LSTM模型在机器翻译中表现优越，并且在不同任务场景下需要考虑选择不同的模型。最后，总结了实验并指出引入注意力机制作为LSTM模型的改进方法。

The article combines the development history of machine translation and focuses on the issue of the loss of contextual information due to dependency during the translation process. Firstly, the LSTM neural network model and its related variants are explored, and a machine translation encoder-decoder framework is established. Secondly, the Pytorch deep learning framework is used to implement the machine translation function from German to English on the Multi30K dataset. Then, using BLEU score as the evaluation metric, the experimental results of LSTM and GRU models were compared. The experiment showed that in machine translation, the LSTM model performed better, and different models should be selected for different task scenarios. Finally, the experiments are summarized and the attention mechanism is proposed as an improvement method for the LSTM model.