In recent years, extending variational autoencoder's framework to learn\ndisentangled representations has received much attention. We address this\nproblem by proposing a framework capable of disentangling class-related and\nclass-independent factors of variation in data. Our framework employs an\nattention mechanism in its latent space in order to improve the process of\nextracting class-related factors from data. We also deal with the multimodality\nof data distribution by utilizing mixture models as learnable prior\ndistributions, as well as incorporating the Bhattacharyya coefficient in the\nobjective function to prevent highly overlapping mixtures. Our model's encoder\nis further trained in a semi-supervised manner, with a small fraction of\nlabeled data, to improve representations' interpretability. Experiments show\nthat our framework disentangles class-related and class-independent factors of\nvariation and learns interpretable features. Moreover, we demonstrate our\nmodel's performance with quantitative and qualitative results on various\ndatasets.\n