Survey of Abstract Meaning Representation: Then, Now, Future

TL;DR

The paper surveys Abstract Meaning Representation (AMR), a graph-based semantic representation that encodes who does what to whom while abstracting away surface syntax. It covers the history and design of AMR, annotation guidelines, and enrichment strategies, then details text-to-AMR parsing and AMR-to-text generation, including evaluation metrics and evolving neural and large-language-model approaches. It also surveys non-English AMR research, multilingual corpora, and cross-lingual parsing/generation, highlighting resources like MASSIVE-AMR and XL-AMR. Finally, it discusses applications across text generation, classification, information extraction, and information seeking, and outlines future directions such as uniform meaning representations and deeper integration with multilingual and multimodal tasks.

Abstract

This paper presents a survey of Abstract Meaning Representation (AMR), a semantic representation framework that captures the meaning of sentences through a graph-based structure. AMR represents sentences as rooted, directed acyclic graphs, where nodes correspond to concepts and edges denote relationships, effectively encoding the meaning of complex sentences. This survey investigates AMR and its extensions, focusing on AMR capabilities. It then explores the parsing (text-to-AMR) and generation (AMR-to-text) tasks by showing traditional, current, and possible futures approaches. It also reviews various applications of AMR including text generation, text classification, and information extraction and information seeking. By analyzing recent developments and challenges in the field, this survey provides insights into future directions for research and the potential impact of AMR on enhancing machine understanding of human language.

Figures (10)

• Figure 1: Abstract Meaning Representation for five different input sentences (A) with the same semantics. (B) shows the AMR with PennMan notation and (C) represents the corresponding graph.
• Figure 2: MathAMR 10.1145/3511808.3557567 for input sentence "Find $x^n+y^n+z^n$ general solution". (A) AMR is generated for the input text, with the formula replaced with a placeholder (PL) for the formula. Then the operator tree representation of the formula (B), is integrated into the AMR, replacing the placeholder in AMR with the root of the operator tree, resulting in MathAMR shown in (C).
• Figure 3: Transition-based and Alignment-based approaches for AMR Parsing. Transition-based methods incrementally generate AMRs through a sequence of actions, often guided by a dependency tree (Left). Alignment-based methods first align text spans to AMR nodes and then build the graph based on these alignments (Right).
• Figure 4: Overview of current AMR parsing approaches. (A) Seq-to-seq models generate a linearized AMR, and (B) Graph prediction directly predicts AMR nodes and edges. Both approaches typically involve pre- and post-processing steps.
• Figure 5: JAMR flanigan-etal-2016-generation approach for text generation from Abstract Meaning Representation. (a) The input AMR graph is processed to remove re-entrancies, yielding a tree structure. (b) This tree is then used by a tree-to-text transducer, which (c) generates the final text output.