Credit Text Similarity Calculation Method Combined with Contrast Learning

The text matching task aims at mining the inherent semantic features from two texts and predicting the correlation or contradiction between texts. The credit text phases have obvious domain-specific features, and there are more adversarial texts, and most of them are of different lengths leading to missing information and unclear main ideas, which will reduce the accuracy of similarity calculation. To this end, this paper proposes a credit text similarity calculation method combined with contrast learning. First, we will recall the candidate set by multi-level text recall method to reduce the influence of non-critical factors and reduce the computational effort of text matching model from shallow semantics, deep semantics and entities. And then fine-tunes the pre-trained model by combining with contrast learning method to calculate the semantic text similarity. The comparison experimental results with BERT and SimCSE show that the text similarity calculation method proposed in this paper works better.