A Simple and Effective Test-Time Augmentation Strategy
Xinchun Zhang, Yu Xiang · 2023
Test-time augmentation (TTA) is a technique that improves model predictive performance by aggregating augmented samples without requiring additional training. While TTA works with any data modality, its use in natural language processing (NLP) is limited. This is partly because the text is complex and inappropriately introducing noise can easily change the semantics and syntax, making it difficult to ensure label consistency. Existing TTA strategies lack generality and are only effective for certain datasets, ignoring the differences between example sentences and word importance. To solve this problem, we introduce interpretable machine learning methods to study the differential effects of enhancers of different importance. We propose a more efficient and robust TTA strategy - STTA, which explicitly assigns an importance to each word for different words in different sentences. We select appropriate enhancement words through measurement, which greatly reduces the labeling problem. Variation risk, significantly improving model performance. To our knowledge, STTA is the first sample difference considered for TTA strategies. A large number of experiments have proven that compared with standard TTA, STTA further improves model performance and provides a feasible and effective TTA paradigm for NLP.