Neural Machine Translation in Electrical Engineering with Cross-layer Information Fusion and Multiple Positional Mapping

This study focuses on the English-Chinese neural machine translation task of professional texts in the field of electrical engineering. Such texts are usually dense in terms, complex in sentence structure, deeply nested in structure, and have low resources, which poses significant challenges to existing neural machine translation models. Aiming at the problem of attenuation of inter-layer semantic information and position information in text encoding of Transformer model in the field of electrical engineering, an optimization method based on cross-layer information fusion and multiple position mapping is proposed. This method integrates the Long Short-Term Memory (LSTM) and the attention mechanism to construct a cross-layer information compensation architecture consisting of a cross-layer feature propagation module and a feature fusion module to compensate for the attenuation of cross-layer semantic information and position information. In addition, in view of the fact that Transformer position information is easily distorted, this study also designed a multiple position mapping module based on dynamic linear transformation, which enhances the robustness of position representation through a multiple superposition mechanism and improves the position sensitivity of the model during deep calculations. Experimental results on translation tasks in the electrical engineering field show that the improved model outperforms the baseline and other comparison models in terms of BLEU, METEOR and TER scores, verifying the effectiveness and superiority of this method in semantic modeling and position information retention in professional field translation tasks.