An Semantic Similarity Matching Method for Chinese Medical Question Text

Liru Wang,Tongxuan Zhang,Jiewen Tian,Hongfei Lin

DOI: https://doi.org/10.1007/978-981-19-9865-2_6

2023-01-01

Abstract:For the purpose of capturing the semantic information accurately and clarifying the user’s questioning intention, this paper proposes a novel, ensemble deep architecture BERT-MSBiLSTM-Attentions (BMA) which uses the Bidirectional Encoder Representations from Transformers (BERT), Multi-layer Siamese Bi-directional Long Short Term Memory (MSBiLSTM) and dual attention mechanism (Attentions) in order to solve the current question semantic similarity matching problem in medical automatic question answering system. In the preprocessing part, we first obtain token-level and sentence-level embedding vectors that contain rich semantic representations of complete sentences. The fusion of more accurate and adequate semantic features obtained through Siamese recurrent network and dual attention network can effectively eliminate the effect of poor matching results due to the presence of certain non-canonical texts or the diversity of their expression ambiguities. To evaluate our model, we splice the dataset of Ping An Healthkonnect disease QA transfer learning competition and “public AI star” challenge - COVID-19 similar sentence judgment competition. Experimental results with CC19 dataset show that BMA network achieves significant performance improvements compared to existing methods.