Neural machine translation (NMT) aims to use computers to translate one language into another, and it plays a critical role in various scientific fields. Since 2014, NMT has developed rapidly, from recursive neural networks [
1], to convolutional neural networks [
2], and then to the Transformer neural network based on self-attention [
3], which has achieved good results. Among several NMT models, Transformer performs the best, both in terms of efficiency and translation quality.
As various scientific fields continue to develop, the demand for NMT is also increasing rapidly. Different professional fields have different professional corpora, and some fields have very limited parallel corpora resources. Traditional NMT cannot meet the translation needs of some professional fields. The English–Chinese corpus in the field of electrical engineering is a typical low-resource corpus. The traditional Transformer does not perform well in the English–Chinese corpus in the field of electrical engineering, often causing mistranslation or misinterpretation of certain feature information in sentences, which makes it difficult for personnel in the electrical industry to use professional equipment and read professional English literature. The field of electrical engineering plays a crucial role in the development of many scientific fields. Therefore, it is essential to study how to design an efficient and stable model on a small-scale parallel corpus to improve the current situation of NMT in the field of electrical engineering.
Low-resource neural machine translation has long been an area of interest in natural language processing, and many researchers have made significant efforts to address this problem. Common improvement methods include data augmentation, introducing prior knowledge, and structural improvements. Tonja used monolingual source-side data to improve low-resource neural machine translation and achieved significant results on the Wolaytta–English corpus, further fine-tuning the best-performing self-learning model which resulted in +1.2 and +0.6 BLEU score improvements for Wolaytta–English and English–Wolaytta translations, respectively [
4]. Mahsuli, MM proposed a method to model the length of a target (translated) sentence given the source sentence using a deep recurrent neural structure—and apply it to the decoder side of neural machine translation systems to generate translation sentences with appropriate lengths which have a better quality [
5]. Pham, NL; Nguyen, V; and Pham, TV used back-translation to enhance the parallel database of English–Vietnamese machine translation, significantly improving the translation quality of the model [
6]. Laskar, SR improved English–Assamese machine translation through pre-training models, and the best MNMT model, Transformer (transliteration-based phrase-augmentation), attained scores of +0.58, +1.86 (BLEU) [
7]. Park, YH enhanced low-resource neural machine translation data through EvalNet and the NMT systems for English–Korean and English–Myanmar, built with the guidance of EvalNet, and achieved 0.1~0.9 gains in BLEU scores [
8]. While these methods have achieved good results, they often require significant time and cost in the data preprocessing stage and have certain drawbacks. Dhar, P introduced bilingual dictionaries to improve Sinhala–English, Tamil–English, and Sinhala–Tamil translation and introduced a weighted mechanism based on small-scale bilingual dictionaries to improve the measurement of semantic similarity between sentences and documents [
9]. Gong, LC achieved good results on several low-resource datasets by guiding self-attention with syntactic graphs [
10]. Hlaing, ZZ added an additional encoder to the transformer model to introduce part-of-speech tagging, improving Thai-to-Myanmar, Myanmar-to-English, and Thai-to-English translation, outperforming such models developed through the existing Thai POS tagger in terms of BLEU scores (+0.13) and chrF scores (+0.47) for Thai-to-Myanmar, and BLEU scores (+0.75) and chrF scores (+0.72) for Myanmar-to-Thai translation pairs [
11]. Considering that convolutional neural networks can extract feature information from sentences, this paper integrates a convolutional neural network as a feature extraction layer into the Transformer model. This method can introduce feature information into the Transformer model without additional processing of the corpus, improving the translation quality of the Transformer model while also saving research time and costs. The main contributions of this article are as follows:
In order to address the issue of feature information misinterpretation and omission in the corpus of electrical engineering with Transformer, a method is proposed to integrate a convolutional neural network as a feature extraction layer into Transformer, which effectively improves the translation accuracy of the model.
Comparative experiments and ablation experiments are designed to verify the performance of the two models proposed in this paper on the dataset of electrical engineering, and their performance is compared with the baseline model, demonstrating that the Transformer model integrated with a convolutional neural network has a better performance.