PropBank

Unlike generalized thematic roles, PropBank defines rolesets that are specific to each verb sense. Each sense of a verb like 'run' or 'break' receives a unique frameset with numbered arguments.

• Arg0: Proto-Agent (causer, experiencer)
• Arg1: Proto-Patient (thing affected, theme)
• Arg2–Arg5: Verb-specific roles (instrument, benefactive, start point, end point)
• ArgM: General modifiers (manner, temporal, locative, negation)

PropBank adds a semantic layer directly onto the syntactic constituency trees of the Penn Treebank. Each predicate is annotated with its frameset, and each argument is linked to a specific constituent node in the parse tree.

This tight coupling of syntax and semantics enables models to learn the mapping between grammatical structure and predicate-argument relations.

While arguments are numbered per verb, PropBank's Arg0 and Arg1 labels exhibit consistent proto-role semantics across the entire corpus.

• Arg0 consistently maps to Dowty's Proto-Agent properties (volition, causation, sentience)
• Arg1 consistently maps to Proto-Patient properties (change of state, incremental theme, affectedness)

This consistency allows models to generalize across unseen predicates.

The original PropBank annotated only newswire text from the Wall Street Journal. The OntoNotes project expanded PropBank-style annotation to multiple genres:

• Newswire and broadcast news
• Broadcast conversation and telephone speech
• Web text and weblogs
• Pivot text (Old and New Testament)

This diversity makes PropBank-trained SRL systems robust across domains.

Beyond verbs, PropBank annotates nominalizations (e.g., 'the destruction of the city') and light verb constructions (e.g., 'take a walk', 'make a decision').

In light verb constructions, the semantic predicate is the noun ('walk', 'decision'), while the light verb ('take', 'make') carries grammatical features. This unified treatment captures meaning regardless of syntactic realization.

PropBank annotations form the core of the CoNLL-2005, CoNLL-2008, CoNLL-2009, and CoNLL-2012 Shared Tasks, which standardized SRL evaluation.

These benchmarks drove the development of neural SRL architectures and established the span-based and dependency-based SRL paradigms. The CoNLL-2012 format remains the de facto standard for SRL model training and evaluation.

The two core sub-tasks of SRL. Argument Identification detects which constituents are arguments of a predicate. Argument Classification assigns the specific role label (e.g., Arg0, Arg1). Modern span-based SRL architectures solve both jointly by scoring all possible text spans.

A sequence labeling encoding that transforms SRL into a token classification problem. Tokens are tagged as Beginning, Inside, or Outside of a semantic argument. This scheme enables BERT-based SRL models to predict roles without relying on pre-computed syntactic parse trees.

A semantic formalism that represents sentence meaning as a rooted, directed graph. Unlike PropBank's flat predicate-argument structures, AMR parsing abstracts away from syntax entirely. It uses PropBank framesets as predicates but encodes 'who is doing what to whom' in a unified semantic network.