Abstract Meaning Representation (AMR)

AMR represents a sentence as a rooted, directed, acyclic graph (DAG) . Unlike a syntactic parse tree, the graph structure abstracts away from word order and function words. The root is the central predicate or concept, and directed edges point from heads to their dependents, representing semantic roles. The acyclic property prevents circular definitions, ensuring a strict hierarchical decomposition of meaning from the main event down to its participants.

A core principle of AMR is that sentences with identical meaning but different syntactic structures map to the same AMR graph. This abstracts away from:

• Voice: 'The dog chased the cat' and 'The cat was chased by the dog' share an identical AMR.
• Word Order: The graph is unordered, focusing purely on semantic relations.
• Function Words: Articles, prepositions, and auxiliary verbs are typically omitted unless they carry semantic weight (e.g., negation).

AMR relies on PropBank to define the semantic roles for predicates. Each verb or event is assigned a frameset, and its arguments are labeled with core roles like:

• :ARG0: Proto-Agent (the doer)
• :ARG1: Proto-Patient (the undergoer)
• :ARG2: Instrument, benefactive, or attribute
• :ARG3: Start point
• :ARG4: End point This provides a standardized, machine-readable vocabulary for 'who is doing what to whom'.

Beyond core PropBank arguments, AMR uses a rich inventory of non-core roles to capture circumstantial information. These include:

• :location, :time, :manner, :purpose
• :duration, :frequency, :condition
• :concession, :cause, :instrument This fine-grained labeling allows the graph to encode the full circumstantial context of an event, which is critical for precise semantic parsing.

AMR normalizes concepts to canonical forms. Named entities are assigned a type (e.g., person, country, organization) and linked to a unique identifier, such as a Wikidata ID. Date and time expressions are normalized to a standard format (e.g., date-entity :year 2024 :month 5). This normalization is essential for entity linking and allows downstream systems to query across different phrasings of the same real-world entity.

A powerful feature of AMR is re-entrancy, where a single node in the graph serves as an argument for multiple predicates. This elegantly captures phenomena like control and shared participants. For example, in 'John wants to run', the node for 'John' is both the :ARG0 of 'want-01' and the :ARG0 of 'run-02'. This graph-based variable binding is a key advantage over tree-based semantic representations.