arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2209.00797cs.CL
收藏

随机文本扰动工作,但并不总是

Random Text Perturbations Work, but not Always

Stony Brook University(石溪大学)

Zhengxiang Wang

摘要:

我们提出了三个大规模的实验二进制文本匹配分类任务中的中文和英文随机文本扰动作为一种数据增强方法的NLP的有效性和推广性进行评估。研究发现,增强对三种神经分类模型的测试集性能既有正面影响,也有负面影响,这取决于模型是否在足够的原始训练样本上进行训练。无论用于增加文本的五个随机文本编辑操作是一起应用还是单独应用,都是如此。我们的研究表明,随机文本扰动的有效性是特定于任务的,通常不是积极的。

英文摘要:

We present three large-scale experiments on binary text matching classification task both in Chinese and English to evaluate the effectiveness and generalizability of random text perturbations as a data augmentation approach for NLP. It is found that the augmentation can bring both negative and positive effects to the test set performance of three neural classification models, depending on whether the models train on enough original training examples. This remains true no matter whether five random text editing operations, used to augment text, are applied together or separately. Our study demonstrates with strong implication that the effectiveness of random text perturbations is task specific and not generally positive.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑