基于强化学习的古今汉语句子对齐研究
Sentence Alignment of Ancient and Modern Chinese Based on Reinforcement Learning
喻快 1邵艳秋 1李炜1
作者信息
- 1. 北京语言大学国家语言资源检测与研究平面媒体中心,北京 100083
- 折叠
摘要
基于深度学习的有监督机器翻译取得了良好的效果,但训练需要大量高质量的对齐语料.对于中文古今翻译场景,高质量的平行语料相对匮乏,这使得语料对齐在该领域具有重要的研究价值和必要性.在传统双语平行语料的句子对齐研究中,传统方法根据双语文本中的长度、词汇、共现文字等特征信息建立综合评判标准来衡量两个句对的相似度.此类方法对句子语义匹配的能力有限,并且在多对多的对齐模式上表现不佳.该文利用具有强大语义能力的预训练语言模型,并基于动态规划算法的强化学习训练目标来整合段落全局信息,进行无监督训练.实验结果证明,使用该方法训练得到的模型性能优于此前获得最好表现的基线模型,特别是在多对多对齐模式下,性能提升显著.
Abstract
Supervised machine translation based on deep learning has achieved good results,which it requires a large amount of high-quality aligned parallel corpora for training.For the Chinese historical and modern translation sce-nario,the relative scarcity of high-quality parallel corpora highlights the significant research value and necessity of text alignment in this field.In traditional research on bilingual sentence alignment,conventional methods use features such as length,vocabulary,and co-occurring characters in bilingual texts to establish a comprehensive eval-uation criterion for measuring the similarity between two sentence pairs.Such methods have limited ability to match sentence semantics and perform poorly in the many-to-many alignment mode.This paper proposes using pre-trained language models with powerful semantic capabilities and reinforcement learning training objectives based on dynamic programming algorithms to integrate paragraph-level global.Experimental results demonstrate that the model trained using the method proposed in this paper outperforms the previously best-performing baseline models,with particularly significant improvements in handling many-to-many alignment patterns.
关键词
双语对齐/预训练语言模型/强化学习/动态规划Key words
bilingual alignment/pre-trained language model/reinforcement learning/dynamic programming引用本文复制引用
出版年
2024
NETL
NSTL
万方数据