This paper presents a comprehensive review and proposed strategy for compressing massive Natural Language Generation (NLG) models focused on enhancing the efficiency of transformer-based models for deployment in real-time online applications. The review examines a range of compression techniques, including pruning, quantization, distillation, and knowledge distillation, explaining their roles in reducing the computational and memory requirements of these models. This reduction is crucial in mitigating deployment costs and latency problems, thus improving user experience. The unique complexities of transformer-based models, such as their sequential nature and attention mechanism, necessitate that the compression strategy applied be task-specific to be effective. The outcomes of this study serve as a valuable guide for NLG model developers wanting to optimize their models for online deployment. The effectiveness of these methods is discussed in terms of numerical performance indices such as latency reduction, memory footprint reduction, and efficiency improvement. By leveraging these strategies, researchers can develop more efficient and cost-effective NLG models.