While there have been advances in Natural Language Processing (NLP), their success is mainly gained by applying a self-attention mechanism into single or multi-modalities. While this approach has brought significant improvements in multiple downstream tasks, it fails to capture the interaction between different entities. Therefore, we propose MM-GATBT, a multimodal graph representation learning model that captures not only the relational semantics within one modality but also the interactions between different modalities. Specifically, the proposed method constructs image-based node embedding which contains relational semantics of entities. Our empirical results show that MM-GATBT achieves state-of-the-art results among all published papers on the MM-IMDb dataset.