Graph routing between capsules

Routing methods in capsule networks often learn a hierarchical relationship for capsules successive layers, but the intra-relation between same layer is less studied, while this key factor semantic understanding text data. Therefore, paper, we introduce new network with graph routing to both relationships, where each are treated as nodes of graph. We investigate strategies yield adjacency and degree matrix three different distances from capsules, propose mechanism those capsules. validate our approach on five classification datasets, findings suggest that combining bottom-up top-down attention performs best. Such an demonstrates generalization capability across datasets. Compared state-of-the-art methods, improvements accuracy datasets used were 0.82, 0.39, 0.07, 1.01, 0.02, respectively.