rolelabeling

Role labeling, also known as semantic role labeling (SRL), is a natural language processing task that aims to identify the semantic structure of a sentence by assigning labels to the arguments of a predicate, typically a verb. The labels describe how different constituents relate to the predicate, indicating roles such as who performs the action, who or what is affected, and other contextual relations. In PropBank-style labeling, arguments often map to slots like A0 (usually the agent) and A1 (the patient), while FrameNet uses frame-specific roles such as Agent, Patient, or Instrument.

A typical SRL system detects predicates, identifies their corresponding arguments, and assigns the appropriate semantic roles to those arguments. Outputs usually consist of one or more labeled argument spans for each predicate. The task can be framed as span-based or dependency-based labeling and often relies on syntactic parses to guide argument identification and role assignment.

Approaches to role labeling have evolved from feature-based methods that used hand-crafted linguistic cues to neural models that leverage large pretrained transformers. Modern SRL systems often perform end-to-end labeling, sometimes jointly with predicate detection, and have been extended to multiple languages and cross-lingual settings. Representations may be sentence- or span-centric, emphasizing the relations between predicates and their arguments.

Key datasets include PropBank, which provides predicate-argument structures for English, and FrameNet, which links predicates to more detailed semantic frames. Evaluation typically uses precision, recall, and F1 metrics, and benchmarks cover both sentence-level accuracy and argument identification quality.

Applications of role labeling span information extraction, question answering, machine translation, summarization, and other tasks requiring an explicit representation of who did what to whom.