Improving Neural Machine Translation in the Field of Electrical Engineering by Using Sentence Backbone Information

Due to the limited availability of corpora in the field of Electrical Engineering and the presence of numerous specialized terms, neural machine translation (NMT) performs poorly in translating the sentence backbone information when it is applied to corpora in the field of Electrical Engineering. In response to this issue, A method to improve NMT by using the sentence backbone information is proposed in this paper. In the proposed method, the source language sentences are used as the input of the Sentence Backbone Information Extraction Model to obtain the sentence backbone information, and then the sentence backbone information are incorporated as an auxiliary during the training process of the NMT model. Furthermore, a module called the Sentence Backbone Information Enhancement Module is introduced. It utilizes the dependency parse trees of the source language sentences to generate the sentence backbone mask matrices. These matrices are then applied to the encoder to force the NMT model to pay more attention to the backbones of sentences. On the English-Chinese parallel corpus in the field of Electrical Engineering, the proposed method in this paper outperforms the Transformer baseline translation model by 1.25 BLEU points. And it outperforms the baseline model in both METEOR and ROUGE-L evaluation metrics. It indicates that the proposed method in this paper can effectively improve translation performance.