Semantic Frames

A lexical unit is a pairing of a word with a specific meaning that evokes a particular semantic frame. A single word can belong to multiple frames depending on its sense.

• The verb 'sell' evokes the COMMERCIAL_TRANSACTION frame, profiling the Seller's perspective.
• The verb 'buy' evokes the same frame but profiles the Buyer's perspective.
• The noun 'price' also evokes COMMERCIAL_TRANSACTION, profiling the Money element.
• This grouping reveals how synonyms and related terms share an underlying conceptual structure.

Frames do not exist in isolation; they are organized in a rich network of inter-frame relations that capture conceptual hierarchies and perspectives.

• Inheritance: A child frame inherits all elements from a parent frame (e.g., REVENGE inherits from REWARDS_AND_PUNISHMENTS).
• Perspectivization: Two frames provide different viewpoints on the same scene (e.g., BUYING and SELLING are perspectives on COMMERCIAL_TRANSACTION).
• Subframe: A frame represents a sub-event within a complex event sequence (e.g., ARREST is a subframe of the CRIMINAL_PROCESS frame).
• Using: A frame presupposes another frame as background (e.g., SPEED uses the MOTION frame).

Each lexical unit has valence patterns that specify how its frame elements are realized syntactically. This maps semantic roles to grammatical functions and phrase types.

• For the verb 'give' in the GIVING frame:
  • Donor → External Argument (Subject, NP)
  • Recipient → Object (NP) or Oblique (PP[to])
  • Theme → Object (NP) or Oblique (PP[of])
• These patterns explain why 'She gave him the book' and 'She gave the book to him' are syntactic alternations of the same frame.

Within a frame, the profile is the specific aspect or participant that a word designates and brings into focus, while the base is the entire background conceptual structure required to understand it.

• 'Hypotenuse' profiles a specific line segment, but its base is the entire RIGHT_TRIANGLE frame.
• 'Widow' profiles a woman, but its base requires the MARRIAGE frame and the knowledge that the husband has died.
• This distinction explains why you cannot understand a word without understanding the full frame it evokes.

The NLP task of detecting predicate-argument structures in text. SRL answers 'who did what to whom, when, where, and how.' Modern approaches use span-based neural architectures or BIO tagging with transformers. SRL is the computational process that identifies which frame elements are instantiated in a sentence, bridging surface text to frame semantics.

Generalized semantic categories describing the relationship between a predicate and its arguments. Common roles include:

• Agent: The deliberate initiator of an action
• Patient: The entity undergoing the action
• Theme: The entity being moved or experienced
• Instrument: The tool used to perform an action
• Experiencer: The entity perceiving or feeling These roles form the theoretical foundation for frame elements.

A rooted, directed graph representation of sentence meaning that abstracts away from syntax. AMR encodes 'who is doing what to whom' as a semantic network of concepts and relations, such as :ARG0 and :ARG1. Unlike frame-based approaches, AMR unifies predicate-argument structures with coreference, negation, and modality into a single graph. AMR parsing is the task of generating these graphs from text.