Abstract Meaning Representation (AMR)

AMR represents sentence meaning as a rooted, directed, acyclic graph (DAG). Nodes represent concepts (words, PropBank framesets, or entities). Edges represent semantic roles (e.g., :ARG0 for agent, :ARG1 for patient). The graph is 'rooted' because the main predicate or focus acts as the top-level node. 'Acyclic' means there are no cycles, ensuring a strict hierarchical semantic structure. This graph structure abstracts away from surface syntax, meaning sentences with different grammatical structures but identical meanings share the same AMR graph.

A core design principle of AMR is syntax-agnosticism. It captures 'who is doing what to whom' without representing syntactic idiosyncrasies like word order, function words (articles, prepositions), or voice (active/passive).

• Active/Passive Neutrality: 'The dog chased the cat' and 'The cat was chased by the dog' map to the same AMR graph.
• Implicit Roles: Semantic roles are explicitly labeled, even if they are implicit in the syntax.
• No Tense/Number: AMR does not typically represent inflectional morphology for tense or number, focusing purely on propositional content.

AMR heavily relies on PropBank, a corpus annotated with verb-specific semantic roles. Predicates in an AMR graph are often PropBank framesets (e.g., chase-01). Each frameset defines a set of numbered core arguments (:ARG0, :ARG1) with consistent semantic meanings.

• :ARG0: Typically the agent or causer.
• :ARG1: Typically the patient or theme.
• :ARG2: Instrument, benefactive, or attribute. This allows AMR to normalize different verbs with similar argument structures and provides a standardized inventory of semantic roles.

AMR normalizes entities and concepts to canonical forms. Named entities are linked to a type and a canonical string.

• Entity Typing: Entities are typed (e.g., person, city, organization).
• Wikification: Entities can be linked to external knowledge bases like Wikipedia (e.g., :wiki "New_York_City").
• Date/Number Normalization: Temporal expressions and quantities are parsed into a structured, normalized format (e.g., :date-entity :year 2023). This normalization is critical for entity linking and downstream reasoning tasks, ensuring that 'NYC' and 'New York City' are treated as the same concept.

AMR uses re-entrancy (a node with multiple incoming edges) to represent concepts that participate in multiple semantic relations. This is crucial for control structures and shared arguments.

• Control Verbs: In 'The boy wants to go', the 'boy' is both the :ARG0 of 'want-01' and the :ARG0 of 'go-01'. The graph shows a single 'boy' node with edges pointing to it from both predicates.
• Relative Clauses: The shared entity in a relative clause is represented via re-entrancy. This mechanism allows AMR to capture complex predicate-argument structures that are difficult to represent in a pure tree structure.

To maintain a rooted graph structure, AMR uses inverse roles (e.g., :ARG0-of, :ARG1-of). Instead of a passive verb pointing to its subject as :ARG1, the subject concept points to the verb with an :ARG1-of edge.

• Passive Construction: 'The cat was chased' becomes (c / cat :ARG1-of (c2 / chase-01)).
• Relational Nouns: 'The destruction of the city' uses :ARG1-of from 'city' to 'destroy-01'. This inversion ensures that the graph remains singly rooted and that the focus of the sentence (the grammatical subject) is often the root node.