最强的数据扩增方法竟然是添加标点符号？

浅浅的花香味﹌ 2022-09-10 10:16 2阅读 0赞

今天的这篇文章源自于EMNLP 2021 Findings，论文标题为[《AEDA: **A**n **E**asier **D**ata **A**ugmentation Technique for Text Classification》][AEDA_ _A_n _E_asier _D_ata _A_ugmentation Technique for Text Classification]。实际上用一句话即可总结全文：**对于文本分类任务来说，在句子中插入一些标点符号是最强的数据扩增方法**

#### AEDA Augmentation ####

读者看到这肯定会想问：添加哪些标点符号？加多少？对于这些问题，原论文给出了非常详细的解答，同时这部分也是论文唯一有价值的地方，其他部分的文字叙述基本是在说一些基础概念，或者前人工作等等

首先，可选的标点符号有六个：\{".", “;”, “?”, “:”, “!”, “,”\}。其次，设添加句子标点的个数为 n n n，则

n ∈ \[ 1 , 1 3 l \] n\\in \[1, \\frac\{1\}\{3\}l\] n∈\[1,31l\]

其中， l l l为句子长度。下面给出几个扩增例子

Original a sad , superior human comedy played out on the back roads of life . Aug 1 a sad , superior human comedy played out on the back roads ; of life ; . Aug 2 a , sad . , superior human ; comedy . played . out on the back roads of life . Aug 3 : a sad ; , superior ! human : comedy , played out ? on the back roads of life . \\begin\{array\}\{cc\} \\hline \\textbf\{Original\} & \\text\{a sad , superior human comedy played out on the back roads of life .\} \\\\ \\hline \\textbf\{Aug 1\} & \\text\{a sad , superior human comedy played out on the back roads ; of life ; .\}\\\\ \\hline \\textbf\{Aug 2\} & \\text\{a , sad . , superior human ; comedy . played . out on the back roads of life .\}\\\\ \\hline \\textbf\{Aug 3\} & \\text\{: a sad ; , superior ! human : comedy , played out ? on the back roads of life .\}\\\\ \\hline \\end\{array\} OriginalAug 1Aug 2Aug 3a sad , superior human comedy played out on the back roads of life .a sad , superior human comedy played out on the back roads ; of life ; .a , sad . , superior human ; comedy . played . out on the back roads of life .: a sad ; , superior ! human : comedy , played out ? on the back roads of life .

光说不练假把式，效果究竟几何呢？原论文做了大量文本分类任务的实验，并且与[EDA][]方法进行了比较，而且有意思的是，AEDA在github上的repo是fork自EDA论文的repo，怎么有种杀鸡取卵的感觉

首先看下面一组图，作者在5个数据集上进行了对比（模型为RNN）

![1bcc36fac9342a839addefa71c309965.png_shadow][]

在BERT上的效果如下表所示，为什么上面都测了5个数据集，而论文中对BERT只展示了2个数据集的结果呢？我大胆猜测是因为在其他数据集上的效果不太好，否则没有理由不把其余数据集的结果贴出来

Model SST2 TREC BERT 91.10 97.00 +EDA 90.99 96.00 +AEDA 91.76 97.20 \\begin\{array\}\{c|cc\} \\text\{Model\} & \\text\{SST2\} & \\text\{TREC\} \\\\ \\hline \\text\{BERT\} & 91.10 & 97.00\\\\ \\hline \\text\{+EDA\} & 90.99 & 96.00\\\\ \\hline \\text\{+AEDA\} & \\pmb\{91.76\} & \\pmb\{97.20\}\\\\ \\end\{array\} ModelBERT\+EDA\+AEDASST291.1090.9991.7691.7691.76TREC97.0096.0097.2097.2097.20

#### Reference ####

*  [AEDA: An Easier Data Augmentation Technique for Text Classification][AEDA_ _A_n _E_asier _D_ata _A_ugmentation Technique for Text Classification]

[AEDA_ _A_n _E_asier _D_ata _A_ugmentation Technique for Text Classification]: https://arxiv.org/pdf/2108.13230.pdf
[EDA]: https://arxiv.org/pdf/1901.11196.pdf
[1bcc36fac9342a839addefa71c309965.png_shadow]: /images/20220829/163fa492f53b46b8ac6a67b4efbb5459.png