摘要
论文使用mask-predict解码扩展CMLMs模型,研究了藏汉神经机器翻译方法。并且针对该模型进行改进。在不同领域藏汉数据集上,经过实验,与非自回归模型NAT和标准的Transformer模型进行比较,在不降低解码速度的情况下,与非自回归模型相比提升了4个BLEU以上;与标准的Transformer模型相比能够达到甚至超过Transformer模型性能,同时解码速度更快。
This study uses mask-predict decoding to extend the CMLMs model and investigates the Tibetan-Chinese neural machine translation method.And improvement is made for this model.After experiments on Tibetan-Chinese datasets in different domains,compared with the non-autoregressive model NAT and the standard Transformer model,the decoding speed is not reduced,and the performance is improved by more than 4 BLEU compared with the non-autoregressive model.Compared with a standard Transformer model can meet or even exceed the performance of the Transformer model,while the decoding faster.
作者
严松思
珠杰
汪超
YAN Songsi;ZHU Jie;WANG Chao(Tibet University School of Information Science and Technology,Lhasa 850000;Tibet Informatization Collaborative Innovation Center Jointly Established by the Province and the Ministry,Lhasa 850000)
出处
《计算机与数字工程》
2023年第2期401-404,410,共5页
Computer & Digital Engineering
基金
国家自然科学基金项目(编号:62066042)
教育部人文社会科学研究项目(编号:21YJCZH059)
2021年西藏自治区高校人文社会科学研究项目(编号:SK2021-24)
西藏大学提升计划项目(编号:ZDTSJH21-07)
西藏大学培育计划项(编号:ZDCZJH21-10)
西藏大学珠峰学科建设计划项目(编号:zf22002001)
西藏大学高水平人才培养计划项目(编号:2020-GSP-S176)资助。
关键词
汉藏神经机器翻译
非自回归模型
掩码预测
Chinese-Tibetan neural machine translation
non-autoregressive model
mask-predict
