Skip to main content

AMR and Other Semantic Graph Formalisms

Overview

This document covers graph-based semantic representation formalisms that go beyond surface syntax to represent meaning. These formalisms produce labeled, directed graphs where nodes represent concepts/entities/events and edges represent semantic relations. They share the property of abstracting away from surface word order and syntactic structure.

AMR (Abstract Meaning Representation)

AMR represents sentence meaning as a rooted, directed, labeled graph. Nodes are concepts (often PropBank framesets or AMR-specific concepts); edges are semantic roles (ARG0, ARG1, etc.) and relations (:mod, :location, :time, etc.).

Structural Mapping

AMR ConceptLayers EquivalentNotes
AMR graphpub.layers.annotation#annotationLayer with kind="graph" and formalism="AMR"The entire AMR graph is one annotation layer.
Concept nodepub.layers.annotation#annotationEach AMR concept is an annotation with label = concept name (e.g., want-01, boy, go-02). ontologyTypeRef can point to the PropBank frameset definition.
Root nodeThe annotation with no incoming headIndexOr explicitly marked via features.
Named entityannotation with label = entity type + features containing :name and :opN valuesAMR represents named entities as type + name structure.
Edge (role)pub.layers.annotation#argumentRefARG0, ARG1, etc. map to argumentRef.role. The child concept maps to argumentRef.annotationId.
Non-core relationsargumentRef with role = :mod, :location, :time, :manner, etc.AMR's non-core relations use the same argumentRef mechanism with different role labels.
Reentrancy (shared nodes)Multiple argumentRef entries pointing to the same annotation UUIDAMR allows a concept to fill multiple roles. Layers handles this by having multiple argumentRef entries with the same annotationId.
Constantsannotation.valueString/number constants (dates, quantities, etc.) stored in value.
Polarity/negationannotation.features.polarity = "-"AMR's :polarity - maps to features.
Alignment to tokensannotation.anchor.tokenRefSequence or annotation.tokenIndexAMR-to-token alignments (ISI aligner output) map to anchoring.

AMR-Specific Concepts

AMR FeatureLayers RepresentationNotes
Multi-sentence AMRMultiple annotation layers (one per sentence) + pub.layers.graph#graphEdgeSet for cross-sentence linksCross-sentence coreference and relations use graph edge sets.
AMR coreferencepub.layers.annotation#clusterSet or AMR reentrancyWithin-sentence: reentrancy (shared UUIDs). Cross-sentence: clusterSet.
Document-level AMRpub.layers.graph#graphEdgeSetRelations between sentence-level AMR graphs.

UCCA (Universal Conceptual Cognitive Annotation)

UCCA represents semantic structure using a directed acyclic graph (DAG) anchored to token spans. Nodes are categorized by semantic type; edges are labeled with foundational semantic categories.

UCCA ConceptLayers EquivalentNotes
UCCA graphannotationLayer(kind="graph") with formalism="UCCA"DAG representation.
Scene (S)annotation with label="Scene" and childrenScene-evoking node.
Participant (A)argumentRef with role="A" (Actor)Participant argument.
Process (P)argumentRef with role="P"Process predicate.
State (S)argumentRef with role="S"State predicate.
Center (C) / Elaborator (E)argumentRef with appropriate role labelsUCCA edge categories map to argument roles.
Remote edgesargumentRef with feature isRemote=trueUCCA's remote edges (reentrancy) are marked via features.
Terminal nodesannotation with anchor.tokenRefSequenceLeaf nodes anchored to text.
Non-terminal nodesannotation with childIdsInternal nodes with children.

DRS (Discourse Representation Structures)

DRS provides box-based semantic representations with discourse referents and conditions.

DRS ConceptLayers EquivalentNotes
DRS boxannotation with kind="graph" and formalism via formalismUriA discourse box. Children are conditions and referents.
Discourse referentannotation with label = variable nameVariables introduced in a box.
Conditionannotation with label = predicatePredicates over referents.
Nested DRS (conditional, negation)annotation.childIds containing sub-box annotationsBox embedding via parent-child structure.
Cross-box anaphoraargumentRef or pub.layers.graph#graphEdgeReferences across box boundaries.

EDS/DMRS (Elementary/Dependency Minimal Recursion Semantics)

EDS and DMRS are graph-based representations derived from HPSG/MRS.

EDS/DMRS ConceptLayers EquivalentNotes
EDS graphannotationLayer(kind="graph") with appropriate formalismDependency-style semantic graph.
Predication nodeannotation with label = predicate symbolEach predication maps to an annotation.
Argument edgeargumentRefARG1, ARG2, etc.
Quantifier scopeargumentRef with role="RSTR" / role="BODY"Scope relations.
Token alignmentannotation.anchor.tokenRefSequenceCharacter-span-to-node alignment.

Semantic Dependency Parsing (SDP)

Formats from the SemEval 2014/2015 shared tasks (DM, PAS, PSD):

SDP ConceptLayers EquivalentNotes
Semantic dependency graphannotationLayer(kind="graph") with appropriate formalismToken-level semantic dependencies.
Predicate nodeannotation with tokenIndex + label (predicate sense)Token identified as predicate.
Argument edgeannotation.headIndex + annotation.labelDirected edge with role label.
Top nodeAnnotation with feature isTop=trueGraph root.