Enhancing AMR-to-Text Generation with Dual Graph Representations

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3183–3194, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3183 Enhancing AMR-to-Text Generation with Dual Graph Representations Leonardo F. R. Ribeiro†, Claire Gardent‡ and Iryna Gurevych† †Research Training Group AIPHES and UKP Lab, Technische Universit¨at Darmstadt www.ukp.tu-darmstadt.de ‡CNRS/LORIA, Nancy, France claire.gardent@loria.fr Abstract Generating text from graph-based data, such as Abstract Meaning Representation (AMR), is a challenging task due to the inherent difficulty in how to properly encode the structure of a graph with labeled edges. To address this dif- ficulty, we propose a novel graph-to-sequence model that encodes different but complemen- tary perspectives of the structural information contained in the AMR graph. The model learns parallel top-down and bottom-up rep- resentations of nodes capturing contrasting views of the graph. We also investigate the use of different node message passing strategies, employing different state-of-the-art graph en- coders to compute node representations based on incoming and outgoing perspectives. In our experiments, we demonstrate that the dual graph representation leads to improvements in AMR-to-text generation, achieving state-of- the-art results on two AMR datasets1. 1 Introduction Abstract Meaning Representation (AMR; Ba- narescu et al. (2013)) is a linguistically-grounded semantic formalism that represents the meaning of a sentence as a rooted directed graph, where nodes are concepts and edges are semantic relations. As AMR abstracts away from surface word strings and syntactic structure producing a language neu- tral representation of meaning, its usage is benefi- cial in many semantic related NLP tasks, includ- ing text summarization (Liao et al., 2018) and ma- chine translation (Song et al., 2019). The purpose of AMR-to-text generation is to produce a text which verbalises the meaning en- coded by an input AMR graph. This is a chal- lenging task as capturing the complex structural information stored in graph-based data is not triv- ial, as these are non-Euclidean structures, which 1Code is available at https://github.com/UKPLab/emnlp2019-dualgraph implies that properties such as global parametriza- tion, vector space structure, or shift-invariance do not hold (Bronstein et al., 2017). Recently, Graph Neural Networks (GNNs) have emerged as a powerful class of methods for learning effec- tive graph latent representations (Xu et al., 2019) and graph-to-sequence models have been applied to the task of AMR-to-text generation (Song et al., 2018; Beck et al., 2018; Damonte and Cohen, 2019; Guo et al., 2019). In this paper, we propose a novel graph-to- sequence approach to AMR-to-text generation, which is inspired by pre-neural generation algo- rithms. These approaches explored alternative (top-down, bottom-up and mixed) traversals of the input graph and showed that a hybrid traver- sal combining both top-down (TD) and bottom-up (BU) information was best as this permits integrat- ing both global constraints top-down from the in- put and local constraints bottom-up from the se- mantic heads (Shieber et al., 1990; Narayan and Gardent, 2012). Similarly, we present an approach where the input graph is represented by two separate struc- tures, …

Content truncated to 500 words