With the accelerated growth of the count of academic papers, the establishment of an effective prediction model to evaluate the academic impact of potential papers has become a hot topic of academic attention, aiming at helping scholars quickly identify high-quality papers and inspire them with the most popular research topics. With the extensive deployment of deep learning, as well as in the field of citation prediction, it has been verified that the metadata text of academic papers has a tremendous impact on the count of citations. In this work, a deep learning model based on Bi-LSTM and Transformer is proposed to further improve the performance of such citationcount prediction model in the field of artificial intelligence, owing to the dedicated engagement of the metadata text from abstract, early 5-years citations and cross-attention mechanisms. The experimental results show that the proposed model can outperform the most recent baselines, indicating its great advantages in the field of citation prediction.