Multidimensional Robustness Evaluation in Text Matching Models

With the rapid development of NLP technology, text-matching models have been widely used in several fields. Still, their vulnerability in the face of small changes in input data or malicious attacks has become a focus of attention. In this study, a comprehensive robustness index (RI) is constructed by multi-dimensional robustness assessment using metrics such as robustness radius, attack success rate (ASR), and performance loss rate (PDR), and the model is extensively perturbed and tested using tools such as TextFlint. The experimental results show that the accuracy of the model decreases on the standard test set. Still, the multi-dimensional robustness assessment effectively reflects the robustness performance of the model in specific application scenarios through the constructed comprehensive robustness index (RI), which provides a thorough assessment framework for identifying model weaknesses and optimizing them.