| 田科,张家俊.基于预训练模型的机器翻译译文检测方法[J].情报工程,2020,6(5):015-026 |
| 基于预训练模型的机器翻译译文检测方法 |
| Machine-Translated Text Detection Method Based on Pre-trained Model |
| |
| DOI：10.3772/j.issn.2095-915X.2020.05.002 |
| 中文关键词: 机器翻译译文；预训练语言模型；双语语料 |
| 英文关键词: Machine-translated text; pre-trained language models; bilingual corpus |
| 基金项目: |
| 作者 | 单位 | | 田科 | 1.中国科学院自动化研究所模式识别国家重点实验室； 2.中国科学院大学人工智能学院 | | 张家俊 | 1.中国科学院自动化研究所模式识别国家重点实验室； 2.中国科学院大学人工智能学院 |
|
| 摘要点击次数: 2085 |
| 全文下载次数: 1783 |
| 中文摘要: |
| 机器翻译译文检测任务旨在大规模文本中判别每句话是机器翻译译文还是人工翻译译文。现有的机器翻译译文检测方法大都采用统计的方法提取特征，但是基于统计的方法提取特征能力有限，严重依赖于离散的手工特征，而神经网络模型使用分布式表示，构建代价较低且能表达细粒度的句法、语义特征差别。在本文中，我们提出使用预训练语言模型和双向门控循环单元模型结合，提取机器翻译译文的语言风格、惯用词等隐层表示作为特征来检测机器翻译译文，检测结果相较之前的统计方法有很大的提升。本文尝试使用所提方法过滤混合机器翻译译文的双语语料，过滤后的语料相较原始的语料规模减小了，但是模型的性能却略有提升。 |
| 英文摘要: |
| The task of machine-translated text detection is to determine whether a sentence is translated by machine or human. Most of the existing detection models use statistical approaches for feature extraction. However, statistical methods have limited feature extraction capabilities and rely heavily on discrete manual features, while neural network models use distributed representations to implicitly express syntactic or semantic features. In this paper, we combine pre-trained language models and bi-directional gated recurrent unit model as a feature extractor. Then implicit features such as language style, idiomatic words are extracted. Experimental results show that our model significantly outperforms previous statistical methods. This paper further uses the proposed method to filter the bilingual corpus where machine-translated texts have been mixed in. The filtered corpus is smaller than the original one, but the performance of the model is slightly improved. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
