Journal of Applied Science and Engineering

 Copyright The Author(s). This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are cited.

Download Citation:  BibTeX | http://dx.doi.org/10.6180/jase.202611_34.063  

Download PDF

In recent years, neural network-based approaches have revolutionized the field of machine translation (MT), with Transformer models and their self-attention mechanisms becoming the dominant paradigm. This research presents an improved neural network approach for English translation, focusing on refining output quality and reducing redundancy through an enhanced self-attention mechanism. The Monarch Butterfly Optimized
Bidirectional Encoder Representations from Transformers with self-attention (MBO-BERT-SAtt) is a model that dynamically weighs contextual information to produce fluent and lexically diverse translations while minimizing repetitive elements. The research utilizes a 2000-pair English-Japanese parallel translation dataset, a large-scale bilingual corpus combining Japanese-English parallel texts. Data preprocessing includes normalization and tokenization of long sentence pairs to improve model training efficiency. The model incorporates a dynamic attention head weighting mechanism, which selectively prunes redundant heads during training to optimize translation performance and computational resources. For evaluation, key metrics such as BLEU, and METEOR are used to quantify translation accuracy, fluency, and redundancy reduction, respectively. The Python implementation of the proposed approach achieves 97.4% recall, 98.5% accuracy, 96.6% F1-score, 70.5 BLEU, 67.6 METEOR, 1.0 s translation time, and 5.4% error rate. Human evaluation further confirms enhanced readability and contextual coherence. This approach demonstrates that refining neural translation models through targeted SAtt mechanisms and redundancy-aware training pipelines leads to more precise, natural, and efficient English translations. The results highlight the potential of combining neural network advancements with linguistic insights for next-generation MT systems capable of handling complex language phenomena such as repetition.

Keywords: Self-Attention Mechanism, Translation Quality Enhancement, Redundancy Reduction, Translation Refinement, English Translation, Bidirectional Encoder Representations from Transformers with self-attention (MBO-BERT-SAtt)