AMR-to-text Generation with Graph Structure Reconstruction and Coverage

Anonymous

Abstract: Generating text from semantic representations such as AMR is a challenging task. Previous research formalizes this task as a graph-to-sequence learning problem and uses various graph neural networks to model the graph structure. Recently, methods based on pre-trained models improve the performance significantly due to pre-trained on a large text corpus. However, these pre-trained model-based methods take linearized AMR graphs as input and may lose the information of graph structure. In addition, these methods don't consider the coverage of the AMR graph. Therefore, some nodes in the graph may be lost or repeated in the generated text. To address these problems, we propose a graph structure and coverage enhanced model for this task. To enhance the information of graph structure, we design two auxiliary objectives, relationship prediction and distance prediction of nodes in AMR graphs. To consider the coverage of AMR graphs, we design a coverage mechanism to solve the problem of information under-translation or over-translation in AMR-to-text generation. Experimental results on three standard datasets show that our proposed method outperforms the existing methods significantly.