Text Matching Model Combining Ranking Information and Negative Example Smoothing Strategies

Aiming at the problems that current text matching methods are difficult to accurately capture the fine-grained ranking information between texts and the insufficient information interaction between different negative examples, a text matching model combining ranking information and negative example smoothing strategy is proposed. Firstly, it ensures the consistency of the ranking of two sentence representations of the input text obtained after different Dropout masks through Jensen-Shannon Divergence. Secondly, it utilizes the pre-trained SimCSE as the teacher model to obtain coarse-grained ranking information and distills this information into the student model through the ListNet sorting algorithm to obtain fine-grained ranking information. Finally, the negative examples are augmented by a negative example smoothing strategy, which effectively solves the problem of insufficient information interaction between negative examples without increasing the batch size. Experimental results on the standard semantic text similarity task show that the proposed model achieves a significant improvement in the Spearman correlation coefficient evaluation metrics compared with existing state-of-the-art methods, proving its effectiveness.

@article{Cai2024,
  title     = {Text Matching Model Combining Ranking Information and Negative Example Smoothing Strategies},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {6},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Xiaodong Cai and Lifang Dong and Yeyang Huang and Mingyao Chen},
  doi       = {10.14569/IJACSA.2024.0150679},
  url       = {https://doi.org/10.14569/IJACSA.2024.0150679}
}