Large-scale text data is characterized by an enormous volume， and the same vocabulary may carry completely different meanings in diverse contexts. Relying solely on fixed rules or models makes it difficult to adapt to dynamic semantic changes， which leads to information loss and semantic incompleteness. In such cases， deep semantic information and contextual relationships cannot be captured， thereby impairing the accuracy of semantic matching. To address this issue， a large-scale text semantic matching method based on an improved bidirectional encoder representations from transformers （BERT） pre-training model is proposed in this paper. Therefore， a large-scale text semantic matching method based on an improved BERT pretrained model is proposed.” The improved BERT pre-training model is applied to enhance the contextual information features of text via positional encoding of text word vectors， thus capturing the contextual information of text effectively. Furthermore， the attention mechanism is adopted to dynamically calculate the feature fusion weights， and a weighted fusion method is used to generate the fused semantic features of text. The semantic consistency between texts to be matched is evaluated through four steps： text feature information extraction， multi-dimensional knowledge encoding， fused semantic tag generation， and semantic matching relationship prediction. A consistency threshold of 0.8 is set， meaning that the texts to be matched are considered to have high semantic consistency when the predicted value exceeds 0.8， thus achieving accurate text semantic matching. Test results show that the Mean Reciprocal Rank （MRR） obtained based on large-scale text sample data is higher than 0.7， and the matching results are more accurate compared with the contrast methods.