The task of graph-to-text generation aims at producing sentences that
preserve the meaning of input graphs. As a crucial defect, the current
state-of-the-art models may mess up or even drop the core structural
information of input graphs when generating outputs. We propose to tackle th