Technical gazette, Vol. 32 No. 3, 2025.
Original scientific paper
https://doi.org/10.17559/TV-20231120001119
An Experimental Study on Improved Sequence-to-Sequence Model in Machine Translation
Yuan-shuai Lan
orcid.org/0009-0001-3289-8519
; School of Electronic Information Engineering, Greely university of China, Chengdu, Sichuan 123, Section 2, Chengjian Avenue, Eastern New Area, Jianyang City, Chengdu City, Sichuan Province, China
*
Chuan Li
; -
Xueqin Meng
; -
Tao Zheng
; -
Mincong Tang
; -
* Corresponding author.
Abstract
This paper presents the N-Seq2Seq model for enhancing machine translation quality and efficiency. The core innovations include streamlined attention mechanisms for focusing on crucial details, word-level tokenization to preserve meaning, text candidate frames for prediction acceleration, and relative positional encoding reinforcing word associations. Comparative analyses on English-Chinese datasets demonstrate approximately 4 BLEU score improvements over baseline Seq2Seq and 2 BLEU gains over Transformer models. Moreover, the N-Seq2Seq model reduces average inference time by 60% and 43% respectively. These techniques improve contextual modeling, reduce non-essential information, and accelerate reasoning. Importantly, the model achieves higher accuracy with low overhead, making it possible to deploy on mobile applications, while the Chinese-centric design can also be quickly adapted to other languages.
Keywords
BLEU evaluation; machine translation; N-Seq2Seq model; Seq2Seq model; WoBert model
Hrčak ID:
330559
URI
Publication date:
1.5.2025.
Visits: 542 *