Attention-based Iterative Decomposition for Tensor Product Representation