Nowadays, people tend to join multiple social networks to enjoy different kinds of services. User identity linkage across social networks is of great importance to cross-domain recommendation, network fusion, criminal behaviour detection, etc. Because of the high cost of manually labeled identity linkages, the semi-supervised methods attract more attention from researchers. Different from previous methods linking identities at the pair-wise sample level, some semi-supervised methods view all identities in a social network as a whole, and align different networks at the distribution level. Sufficient experiments show that these distribution-level methods significantly outperform the sample-level methods. However, they still face challenges in extracting features and processing sample-level information. This paper proposes a novel semi-supervised framework with efficient feature extraction and network alignment for user identity linkage. The feature extraction model learns node embeddings from the topology space and feature space simultaneously with the help of dynamic hypergraph neural network. Then, these node embeddings are fed to the network alignment model, a Wasserstein generative adversarial network with a new sampling strategy, to produce candidate identity pairs. The proposed framework is evaluated on real social network data, and the results demonstrate its superiority over the state-of-the-art methods.