Corpora provide convenient resources for various linguistic analyses. By word alignment of large-scale parallel corpora, high-quality translation models can be built, so that machine translation systems can generate translations more in line with human language habits. In view of the problem that traditional algorithms are difficult to capture such long-distance dependencies effectively, resulting in inaccurate alignment, this paper conducts research based on RNN. Both encoders and decoders use GRUs as the basic cyclic unit to effectively capture contextual information and semantic dependencies in text. The introduction of AM allows the decoder to dynamically focus on different parts of the input sequence while generating each output word, thus better capturing semantic dependencies. By comparing with IBM model, GIZA++ model and Rule-M model, the RNN word alignment algorithm has many advantages.