Abstract: Parallelism is a popular rhetorical device featured by its ability to enhance momentum, emphasize key points, and provide clear structure, and there is now no dedicated generation model or publicly available corpus for parallelism. This paper proposes a Chinese parallelism generation model named CPG-PosDep, developed based on the collection and construction of a parallelism dataset and considering the linguistic characteristics of parallelism. Followinging a linguistic approach, the model employs predefined segmentation markers, intra-sentence word order indicators, and inter-sentence position symbols, combined with a random sampling strategy to globally annotate parallelism. It integrates an enhanced Transformer attention mechanism to learn part-of-speech alignment features, and adopts BERT and attention mechanisms to incorporate dependency relation information from sentences into the model. Experiments indicate that the model can generate coherent sentences that maintain consistent part-of-speech and dependency relations with the given clauses, much better than those by existing couplet or poetry models.