Scholarlink

This article aims to solve the problem of scarce data from a single corpus and difficulty in learning the correspondence between the source and target languages. It uses a multi-translation parallel corpus for neural machine translation (NMT) research: extracting data from multiple parallel corpora and cleaning them, constructing a NMT model based on the transformer architecture. It uses Xavier for parameter initialization and unfolds model training through backpropagation and stochastic gradient descent algorithms. This article adopts the bagging method to integrate different parameter models and optimizes translation results based on language models and phrase tables. The research results indicate that the Bilateral Evaluation Understudy (BLEU) of the paper’s model in the Europarl corpus is 0.84, which is approximately 5.0% higher than the Generative Pre-trained Transformer (GPT). The improvement method adopted can achieve more accurate NMT in data scarcity scenarios.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research on Optimization of Machine Translation Performance Based on Deep Learning Algorithm

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yan Meng,
  • Jiuquan Zhang

摘要

This article aims to solve the problem of scarce data from a single corpus and difficulty in learning the correspondence between the source and target languages. It uses a multi-translation parallel corpus for neural machine translation (NMT) research: extracting data from multiple parallel corpora and cleaning them, constructing a NMT model based on the transformer architecture. It uses Xavier for parameter initialization and unfolds model training through backpropagation and stochastic gradient descent algorithms. This article adopts the bagging method to integrate different parameter models and optimizes translation results based on language models and phrase tables. The research results indicate that the Bilateral Evaluation Understudy (BLEU) of the paper’s model in the Europarl corpus is 0.84, which is approximately 5.0% higher than the Generative Pre-trained Transformer (GPT). The improvement method adopted can achieve more accurate NMT in data scarcity scenarios.