“…Due to the poor encoder effect of the NMT model, Tan, Z. et al proposed the lattice-to-sequence attentional NMT model, which generalizes the CNN encoder to the lattice topology, and the proposed encoder reduces the negative impact of errors in the model operation, and outputs more expressive and flexible translated sentences [17]. Chua, C. C. et al introduced a structural, semantic approach to the EBMT model based on the introduction of a structural semantics approach to optimize EBMT from the meaning level, and the structural semantics ensures the consistency and completeness of sentences in the input process, thus improving the accuracy of translation [18]. Zhao, Y. et al in order to enhance the impact of semantic information in the visual feature capture, proposed the MNMT approach with semantic image regions, utilized the CNNs to improve the performance of the model, constructed the training dataset of 30k size, the proposed algorithm was tested in simulation experiments, and the experimental results show that the proposed algorithm can extract semantic information that matches the translated text according to the visual features and improve the performance of translation [19].…”