Aiming at the problem that the current general-purpose semantic text similarity calculation methods are difficult to use the semantic information of scientific academic conference data, a semantic similarity calculation algorithm for scientific academic conferences by fusion with domain features is proposed. First, the domain feature information of the conference is obtained through entity recognition and keyword extraction, and it is input into the BERT network as a feature and the conference information. The structure of the Siamese network is used to solve the anisotropy problem of BERT. The output of the network is pooled and normalized, and finally the cosine similarity is used to calculate the similarity between the two sessions. Experimental results show that the SBFD algorithm has achieved good results on different data sets, and the Spearman correlation coefficient has a certain improvement compared with the comparison algorithm.