提出一种从整体到局部优化的风格迁移(global-local based style transfer,G-LST)模型.首先,利用广泛的源端数据进行迭代优化来自动构建高质量的伪平行数据,并通过联合训练来提升模型对整体风格的语义感知;随后,利用常识性知识修正词级的细粒度风格来增强局部风格的表现,同时兼顾整体与局部风格,提高风格转换的准确度.基于GYAFC数据集的实验结果表明,相较于目前表现最佳的文本风格迁移模型,G-LST模型在E&M与F&R两个领域数据上的风格转换准确率分别提高了 2.70%和 4.47%,内容保留与风格准确率的综合指标分别提升了 1.18%和 1.95%.
A text style transfer model from global to local optimization
A global-local based style transfer(G-LST)framework is proposed,optimizing from a glob-al to a local level.Firstly,extensive source-side data is used for iterative optimization to automatically construct high-quality pseudo-parallel data,and through joint training to improve the model's seman-tic perception of the global style.Subsequently,the model enhances the local style representation by modifying the style at the word level with commonsense knowledge.This method considers both global and local styles simultaneously,thereby improving the accuracy of style transfer.Experimental results on the GYAFC dataset show that compared with the state-of-the-art text style transfer model,the G-LST model's style transfer accuracy on data in the E&M and F&R fields has increased by 2.70%and 4.47%respectively.The comprehensive metrics for content preservation and style accuracy have improved by 1.18%and 1.95%respectively.
text style transferiterative optimizationjoint trainingcommonsense knowledge
国家科技期刊平台
NSTL
万方数据
