Abstract Meaning Representation (AMR)

AMR represents a sentence as a single connected graph with a top-level root node, typically the main predicate. Directed edges labeled with semantic relations (e.g., :ARG0, :ARG1) connect concepts. This structure naturally handles re-entrancy, allowing a single concept node to participate in multiple relations—something tree-based representations cannot express.

• Root node: The central predicate or focus of the sentence
• Directed edges: Labeled with PropBank-style semantic roles
• Re-entrancy: Shared nodes capture coreference and control structures
• Acyclicity: No circular paths, ensuring a well-formed semantic hierarchy

AMR deliberately discards syntactic idiosyncrasies such as word order, tense, articles, and passive voice markers. Sentences with identical meaning but different syntactic realizations map to the same AMR graph. For example, 'The cat chased the mouse' and 'The mouse was chased by the cat' produce identical AMRs.

• No function words: Articles, prepositions, and auxiliaries are omitted unless semantically contentful
• No word order: Graph topology replaces linear sequence
• Canonicalization: Active/passive alternations normalize to the same structure
• Focus on propositional content: Only meaning-bearing elements are represented

AMR uses a verb-specific semantic role inventory derived from PropBank. Core arguments are labeled :ARG0 through :ARG5, where :ARG0 typically corresponds to the prototypical agent and :ARG1 to the prototypical patient. Non-core modifiers use a fixed set of general semantic relations.

• :ARG0: Prototypical agent or causer
• :ARG1: Prototypical patient or theme
• :ARG2 through :ARG5: Verb-specific roles (e.g., instrument, benefactive)
• General modifiers: :time, :location, :manner, :purpose, :condition
• :polarity: Negation marker
• :mode: Expresses modality (e.g., expressive, interrogative)

AMR normalizes content words to canonical concept identifiers. Nouns, verbs, and adjectives are mapped to their lemma forms or ontological entries. Named entities are typed (e.g., person, country, organization) and linked to external knowledge bases. Dates, numbers, and quantities are normalized to a standard format.

• Lemmatization: 'ran' → run-01
• Named entity typing: 'Barack Obama' → (p / person :name (n / name :op1 "Barack" :op2 "Obama"))
• Wikification: Optional :wiki links to Wikipedia entries
• Quantity normalization: 'three hundred' → 300
• Date normalization: 'July 4, 2024' → (d / date-entity :year 2024 :month 7 :day 4)

Each concept node in an AMR graph is assigned a unique variable (e.g., c for 'cat', m for 'mouse'). These variables enable explicit coreference resolution. When the same variable appears in multiple argument positions, it indicates that the same entity participates in multiple relations—a mechanism called re-entrancy.

• Variable assignment: Every concept receives a distinct identifier
• Re-entrancy example: 'The boy wants to go' — the boy is both the wanter (:ARG0 of want-01) and the goer (:ARG0 of go-01)
• Implicit arguments: Variables can fill roles without overt syntactic realization
• Cross-sentence coreference: Extensions of AMR support document-level entity tracking

AMR defines a set of inverse relations (prefixed with :ARG0-of, :ARG1-of, etc.) to maintain a rooted, directed graph while allowing traversal in both directions. These inverses enable the representation of relative clauses, passive constructions, and other structures where the semantic head differs from the syntactic head.

• :ARG0-of: Inverse of :ARG0 — 'the chased mouse' where mouse is the patient
• :ARG1-of: Inverse of :ARG1 — 'the chasing cat' where cat is the agent
• :mod-of: Inverse of :mod — for nouns modified by relative clauses
• :domain-of: Inverse of :domain — for part-whole relationships
• Preserves rootedness: Ensures every node is reachable from the root via directed paths

The computational task of automatically transforming a natural language sentence into its corresponding AMR graph. Modern approaches use sequence-to-graph transformers or transition-based parsers that predict nodes and edges incrementally.

• Input: Raw text (e.g., "The developer deployed the model")
• Output: A rooted, directed acyclic graph with concept nodes and labeled edges
• Key challenge: Reentrancy prediction—identifying when a concept participates in multiple relations
• State-of-the-art: Fine-tuned encoder-decoder models achieving Smatch scores above 85 on newswire text

A precursor and complementary task to AMR that identifies predicate-argument structures in text. While SRL answers 'who did what to whom,' AMR goes further by normalizing predicates, abstracting away from syntax, and representing coreference explicitly.

• PropBank-style SRL assigns verb-specific roles (Arg0, Arg1) to constituents
• FrameNet-style SRL maps sentences to conceptual frames with frame elements
• AMR unifies both approaches into a single graph representation
• SRL outputs often serve as intermediate features for AMR parsing pipelines

The broader task of converting natural language into a formal meaning representation that machines can execute or reason over. AMR is one target formalism among several:

• λ-calculus expressions: Used for compositional question answering
• SQL queries: For text-to-database translation
• Executable logical forms: For virtual assistants and robotic commands
• AMR: Favored when the goal is language-neutral meaning abstraction rather than execution

AMR distinguishes itself by being syntax-agnostic—sentences with different syntactic structures but identical meaning map to the same graph.

The proposition bank that provides the predicate sense inventory and role definitions underlying AMR's predicate-argument structure. Every AMR predicate node links to a PropBank frameset.

• Contains over 6,000 verb framesets with sense-specific role descriptions
• Defines numbered arguments (Arg0 = prototypical agent, Arg1 = prototypical patient)
• AMR enriches PropBank roles with fine-grained relations like :location, :time, :manner
• PropBank's verb-specific roles are generalized in AMR through the normalization of predicate senses

A lexical database organized around semantic frames—conceptual structures describing situations, objects, or events. While AMR uses PropBank predicates, FrameNet provides an alternative, more cognitively motivated inventory of meaning structures.

• Frame elements define participant roles (e.g., Buyer, Seller, Goods in the Commerce_buy frame)
• Frames capture world knowledge about typical scenarios
• Cross-lingual AMR efforts often leverage FrameNet's language-neutral frame definitions
• AMR-to-FrameNet alignment enables richer downstream reasoning about event structures

The task of identifying all expressions that refer to the same entity in a text. AMR represents coreference explicitly through reentrant nodes—a single concept node with multiple incoming edges.

• Example: "The model was deployed. It performed well." → Both sentences share the same model node
• AMR graphs collapse mentions across sentence boundaries, unlike most SRL systems
• Coreference resolution is a critical preprocessing step for document-level AMR parsing
• The OntoNotes corpus provides gold-standard coreference annotations used in AMR training pipelines