Multi-aware coreference relation network for visual dialog

As a challenging cross-media task, visual dialog assesses whether an AI agent can converse in human language based on its understanding of visual content. So the critical issue is to pay attention not only to the problem of coreference in vision, but also to the problem of coreference in and between vision and language. In this paper, we propose the multi-aware coreference relation network (MACR-Net) to solve it from both textual and visual perspectives and to do fusion in complementary awareness. Specifically, its textual coreference relation module identifies textual coreference relations based on multi-aware textual representation from textual view. Furthermore, the visual coreference relation module adaptively adjusts visual coreference relations based on contextual-aware relations representation from visual view. Finally, the multi-modals fusion module fuses multi-aware relations to get an aligned representation. Extensive experiments on the VisDial v1.0 benchmarks show that MACR-Net achieves state-of-the-art performance.