This website requires JavaScript.

What Do Compressed Multilingual Machine Translation Models Forget?

Alireza MohammadshahiVassilina NikoulinaAlexandre BerardCaroline BrunJames HendersonLaurent Besacier
May 2022
摘要
Recently, very large pre-trained models achieve state-of-the-art results invarious natural language processing (NLP) tasks, but their size makes it morechallenging to apply them in resource-constrained environments. Compressiontechniques allow to drastically reduce the size of the model and therefore itsinference time with negligible impact on top-tier metrics. However, the generalperformance hides a drastic performance drop on under-represented features,which could result in the amplification of biases encoded by the model. In thiswork, we analyze the impacts of compression methods on Multilingual NeuralMachine Translation models (MNMT) for various language groups and semanticfeatures by extensive analysis of compressed models on different NMTbenchmarks, e.g. FLORES-101, MT-Gender, and DiBiMT. Our experiments show thatthe performance of under-represented languages drops significantly, while theaverage BLEU metric slightly decreases. Interestingly, the removal of noisymemorization with the compression leads to a significant improvement for somemedium-resource languages. Finally, we demonstrate that the compressionamplifies intrinsic gender and semantic biases, even in high-resourcelanguages.
展开全部
图表提取

暂无人提供速读十问回答

论文十问由沈向洋博士提出,鼓励大家带着这十个问题去阅读论文,用有用的信息构建认知模型。写出自己的十问回答,还有机会在当前页面展示哦。

Q1论文试图解决什么问题?
Q2这是否是一个新的问题?
Q3这篇文章要验证一个什么科学假设?
0
被引用
笔记
问答