Chinese semantic matching method based on twinborn interaction and fine tuning representation

The invention discloses a Chinese semantic matching method based on twinborn interaction and fine tuning representation, which comprises the following steps of: firstly, completing vector initialization of a text by using a RoBERTa-WWM-EXT pre-training model, and constructing a twinborn structure embedded with a soft alignment attention mechanism (SA-attention) and a BiLSTM training layer aiming at an initial feature vector so as to enhance the semantic interaction between sentence pairs; secondly, two texts to be matched are connected to be connected into a RoBERTa-WWM-EXT pre-training model for vectorization, and a connected vectorization result is input into an LSTM-BiLSTM network layer for enhancement training so as to strengthen the upper and lower semantic relation in a sentence; and then a training model capable of finely tuning RoBERTa-WWM-EXT initial vectors is built to generate text vectors subjected to tag supervision fine tuning, so that the expression strength of the vectors on the semantic relat