15 papers with code • 5 benchmarks • 5 datasets
Abstract Meaning Representation (AMR) is a directed graph of labeled concepts and relations that captures sentence semantics. The propositional meaning behind its concepts abstracts away lexical properties. AMR is tree-like in structure as it has a single root node and few children with multiple parents. The goal of AMR-to-Text Generation is to recover the original sentence realization given an AMR. This task can be seen as the reverse of the structured prediction found in AMR parsing.
We show that the PLMs BART and T5 achieve new state-of-the-art results and that task-adaptive pretraining strategies improve their performance even further.
AMR-to-text generation is a problem recently introduced to the NLP community, in which the goal is to generate sentences from Abstract Meaning Representation (AMR) graphs.
Recent studies on AMR-to-text generation often formalize the task as a sequence-to-sequence (seq2seq) learning problem by converting an Abstract Meaning Representation (AMR) graph into a word sequence.
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.
Our generated text has a significantly better semantic fidelity than the state of the art across all four datasets