Considering the gap between training and inference phases of author imitation models trained with Denoising AutoEncoder loss, this paper proposes a noise generation algorithm for test set based on part-of-speech retention. Leveraging linguistic definitions to identify the part-of-speech associated with the primary elements of a sentence, this algorithm selectively retains elements with corresponding parts-of-speech during the addition of noise, while introducing normal noise to other components of the sentence. The proposed algorithm ensures maximum extraction of source text information during the model's inference stage and reduces the gap between the training and inference stages. Moreover, this paper also improves the noise generation algorithm employed during the training phase of De-noising AutoEncoder. Experimental results demonstrate that the proposed noise generation algorithm achieves a relatively well-balanced combination of stylistic intensity, content preservation, and fluency, surpassing other types of noise generation algorithms.