Investigating the Effect of Relative Positional Embeddings on AMR-to-Text Generation with Structural Adapters

TL;DR

This paper explores how relative position embeddings affect AMR-to-text generation and the robustness of StructAdapt, revealing that RPE may encode input graph information and suggesting further research in this area.

Contribution

It investigates the impact of relative position embeddings on AMR-to-text generation and assesses the robustness of StructAdapt, providing insights into their roles in graph-to-text models.

Findings

RPE may partially encode input graph information

StructAdapt's robustness is influenced by RPE effects

Further research needed on RPE's role in Graph-to-Text tasks

Abstract

Text generation from Abstract Meaning Representation (AMR) has substantially benefited from the popularized Pretrained Language Models (PLMs). Myriad approaches have linearized the input graph as a sequence of tokens to fit the PLM tokenization requirements. Nevertheless, this transformation jeopardizes the structural integrity of the graph and is therefore detrimental to its resulting representation. To overcome this issue, Ribeiro et al. have recently proposed StructAdapt, a structure-aware adapter which injects the input graph connectivity within PLMs using Graph Neural Networks (GNNs). In this paper, we investigate the influence of Relative Position Embeddings (RPE) on AMR-to-Text, and, in parallel, we examine the robustness of StructAdapt. Through ablation studies, graph attack and link prediction, we reveal that RPE might be partially encoding input graphs. We suggest further research regarding the role of RPE will provide valuable insights for Graph-to-Text generation.