STMAP: A novel semantic text matching model augmented with embedding perturbations

Abstract

Semantic text matching models have achieved outstanding performance, but traditional methods may not solve Few-shot learning problems and data augmentation techniques could suffer from semantic deviation. To solve this problem, we propose STMAP, which is implemented from the perspective of data augmentation based on Gaussian noise and Noise Mask signal. We also employ an adaptive optimization network to dynamically optimize the several training targets generated by data augmentation. We evaluated our model on four English datasets: MRPC, SciTail, SICK, and RTE, with achieved scores of 90.3%, 94.2%, 88.9%, and 68.8%, respectively. Our model obtained state-of-the-art (SOTA) results on three of the English datasets. Furthermore, we assessed our approach on three Chinese datasets, and achieved an average improvement of 1.3% over the baseline model. Additionally, in the Few-shot learning experiment, our model outperformed the baseline performance by 5%, especially when the data volume was reduced by around 0.4. Our ablation experiments further validated the effectiveness of STMAP.22We have released our source code. https://github.com/wangyanhao0517/STMAP.