Semantic Role Labeling
Identifying who did what to whom, when, and where — labeling the predicate-argument structure of sentences with semantic roles drawn from resources like PropBank and FrameNet.
Definition
Semantic role labeling is the task of automatically identifying the predicates in a sentence and assigning semantic roles to their arguments.
Scope
Covers the shallow-semantic task of detecting predicates and labeling their arguments with semantic roles (agent, patient, instrument, and the like). It includes the role inventories of FrameNet and PropBank, statistical and neural labeling methods, and evaluation. Full logical-form construction is covered in compositional semantics.
Core questions
- How are predicates and their arguments identified in a sentence?
- What role inventories do FrameNet and PropBank provide?
- How do syntactic features support role labeling?
- How is semantic role labeling used in downstream understanding tasks?
Key concepts
- predicate
- semantic role
- agent and patient
- PropBank
- FrameNet
- argument identification
- thematic role
- shallow semantic parsing
Key theories
- Statistical semantic role labeling
- Gildea and Jurafsky's framing of role labeling as a supervised classification problem over syntactic constituents, using features derived from parse trees.
- Frame- and role-based resources
- Annotated inventories such as FrameNet and PropBank that define predicates and their roles, providing the supervision needed to train role labelers.
History
Building on Fillmore's case grammar and frame semantics, Gildea and Jurafsky's 2002 work made automatic semantic role labeling a standard task once FrameNet and later PropBank supplied annotated data. The task became a fixture of shared evaluations and was later improved by neural models.
Debates
- FrameNet versus PropBank role schemes
- Whether fine-grained frame-specific roles or coarser predicate-specific roles are more useful; each supports different applications and trades coverage against consistency.
Key figures
- Daniel Gildea
- Daniel Jurafsky
- Martha Palmer
- Charles Fillmore
Related topics
Seminal works
- gildea2002
- palmer2010
Frequently asked questions
- How is semantic role labeling different from parsing?
- Parsing recovers grammatical structure, such as which phrase is the subject. Role labeling recovers meaning relations, such as which participant is the agent, which need not coincide with grammatical position.