Personalized recommendation systems offer rapid information access, especially for online news platforms. Generative Adversarial Network (GAN) models have been successfully adopted for recommendation. However, very few GAN-based recommendation systems consider the content of news articles, despite its essential latent features, which may affect a recommendation system's effectiveness. This study proposes a novel hybrid GAN recommendation model, which combines Long Short-Term Memory (LSTM), Matrix Factorization (MF) and latent feature extraction of textual content in the GAN structure. Existing GAN-recommendation combines MF to analyze users' global preferences and LSTM to capture users' dynamic preferences over time, but it does not consider the relative importance of MF and LSTM. To effectively combine LSTM and MF, this study proposes a novel attention mechanism for adjusting MF and LSTM weights by learning their importance. In addition, existing GAN-based recommendation does not consider item text content. Our proposed model improves the GAN model by combining the Collaborative Topic Modeling (CTM) and Convolutional Neural Network-Matrix Factorization (CNN-MF) in the Generator to enhance content feature extraction when deriving user and item latent vectors. Specifically, CTM is adopted to obtain the initial latent user/item features of the GAN model, and CNN-MF is used to enhance the extraction of potential text content features in the Generator. The proposed model can enhance existing GAN-recommendation, and increase the performance of preference predictions on textual content such as news articles. This study conducted experimental evaluations using the dataset collected from a news website. The experiment result shows that the proposed approach outperforms several baseline methods on realworld news recommendation.