In existing methods for entity-relation extraction, both entity-driven and relation-driven approaches commonly suffer from insufficient interaction between entities and relationships. Specifically, there is a lack of utilization of the semantic information inherent in relationships. This paper proposed a relation semantic fusion-based entity relation extraction method (RSFnet). Firstly, all possible subjects are extracted from the sentence, and a mapping mechanism is used to obtain corresponding potential relations. At the same time, we treated relations as prior knowledge and used attention mechanisms to obtain sentence representations with relation semantics. The subject information is used as prior features, and the subject features are obtained through a bi-directional long-short term memory (BiLSTM) network. The updated sentence representations and enhanced subject features are further utilized for object and relation extraction, ultimately outputting triplets. The performance of the proposed model was validated through experimental results on three datasets. Additionally, this paper adopts convolutional encoding, resulting in better inference performance than methods based on Bidirectional Encoder Representations from Transformers (BERT), indicating that our model can improve triplet extraction performance while maintaining inference speed.