Link prediction in dynamic signed networks aims to uncover potential relationships between nodes through known network topologies and attribute features. Most of the current mainstream link prediction methods are designed based on graph representation learning, however, they often fail to simultaneously learn the signed semantic information and temporal information contained in the network. In addition, existing graph neural network representation learning models are also difficult to capture the long-distance information dependence between nodes. To address the above problems, this study proposes a link prediction method based on representation learning (RLLP) in the dynamic signed social network. This method designs a temporal random walk strategy capable of sampling long-distance semantic paths in networks and embeds the complex information carried by each path through balance theory and time-aware long short-term memory neural networks (T-LSTM). Finally, the graph attention mechanism is introduced to generate low-dimensional dense vector representations for nodes, enhancing the predictive capability of network links. The proposed method is verified on three real-world datasets, and the experimental results show that, compared with other baselines, RLLP achieves better performance in both F1-score and accuracy metrics, demonstrating high feasibility and effectiveness in the applications of social network link prediction.