Abstract:
Test-time augmentation (TTA) is a well-established technique that involves aggregating transformed examples of test inputs during the inference stage. The goal is to enhance model performance and reduce the uncertainty of predictions. Despite its advantages of not requiring additional training or hyperparameter tuning, and being applicable to any existing model, TTA is still in its early stages in the field of NLP. This is partly due to the difficulty of discerning the contribution of different transformed samples, which can negatively impact predictions. In order to address these issues, we propose Selective Test-Time Augmentation, called STTA, which aims to select the most beneficial transformed samples for aggregation by identifying reliable samples. Furthermore, we analyze and empirically verify why TTA is sensitive to some text data augmentation methods and reveal why some data augmentation methods lead to erroneous predictions. Through extensive experiments, we demonstrate that STTA is a simple and effective method that can produce promising results in various text classification tasks.
摘要:
测试时间增强(TTA)是一种完善的技术,涉及在推理阶段汇总测试输入的转换示例。目标是增强模型性能并减少预测的不确定性。尽管它的优点是不需要额外的训练或超参数调整,并且适用于任何现有模型,TTA在NLP领域仍处于早期阶段。这部分是由于难以辨别不同转化样品的贡献,这会对预测产生负面影响。为了解决这些问题,我们提出了选择性测试时间增强,叫做STTA,其旨在通过识别可靠的样品来选择用于聚集的最有益的转化样品。此外,我们分析并实证验证了为什么TTA对某些文本数据增强方法敏感,并揭示了为什么某些数据增强方法会导致错误的预测。通过广泛的实验,我们证明STTA是一种简单有效的方法,可以在各种文本分类任务中产生有希望的结果。