面向科技文献神经机器翻译词汇表优化研究

Vocabulary Optimization of Neural Machine Translation for Scientific and Technical Document

【目的】针对面向科技文献的神经机器翻译中存在的词汇表受限问题,提出优化方法,进而提升翻译质量。【方法】根据科技词汇构词规律,结合点互信息,在保留词汇义素完整的同时,对神经机器翻译词汇表进行优化,达到减少未登录词的目的。【结果】选择NTCIR-2010专利语料和自动化计算机领域期刊论文摘要语料进行实验,将实验结果与普通分词和子词分词对比,证明该方法的有效性。【局限】仅考虑中文字符的优化。【结论】在中文科技文献领域,基于科技词汇构词的词汇表优化方法能够提升翻译效果。

[Objective]This paper optimizes the vocabulary of Neural Machine Translation(NMT)in scientific and technical domain for the problem of vocabulary limitation and improves the translation performance.[Methods]Based on the word formation and Point-wise Mutual Information(PMI),the paper proposes a method to optimize the vocabulary while preserving the integrity of the lexical semanteme which reduces the number of unknown words.[Results]The NTCIR-2010 corpus and abstract of journal articles in the domain of automation and computer were selected for experiments.The experimental results were compared with the segmentation method and the sub-word method,and it proved the effectiveness of the method.[Limitations]This paper did not cover the optimization of non-Chinese characters.[Conclusions]The experiments show that in scientific and technical domain,the vocabulary optimization algorithm based on scientific word formation achieves better translation performance.