BERT-Based Mixed Question Answering Matching Model

Abstract

Compared with the current question answering system, similarity matching methods are largely separated into two categories: deep learning methods and conventional ways. Conventional ways rely heavily on artificial features, have weak generalization ability, and insufficient accuracy. RNN and CNN also have text global feature extraction. Limitations. This paper proposes a BERT-based hybrid question answering matching model, which uses the BERT -base pre-training model to capture and represent the semantic information of the QAS sentence and the semantic relevance between the two. The feature vector generated by the BERT model is used as Bi-LSTM _ GCN for the input of the model, feature extraction is performed to further obtain the syntactic features of the sentence, and finally the attention mechanism is added to find the target answer, and the effectiveness of the proposed algorithm is verified on the two types of data sets.