PropBank
PropBank, an abbreviation of Proposition Bank, is an annotated linguistic corpus that represents the predicate–argument structure of sentences. Its principal English release adds semantic information to the Penn Treebank, whose texts had already been annotated for phrase structure and grammatical categories. PropBank associates each occurrence of a predicate with a sense-specific roleset and identifies the constituents that realize its semantic arguments.
The resource was developed for research in computational linguistics, particularly the training and evaluation of systems for semantic role labeling. Its annotation scheme occupies an intermediate level between syntactic analysis and fully specified semantic representation. It records who or what participates in an event, while generally leaving reference, temporal interpretation, logical scope, and discourse relations to other forms of annotation.
Development
PropBank originated as an extension of the Penn Treebank's approximately one-million-word collection of The Wall Street Journal text. The project was directed by Martha Palmer, with major contributions from Daniel Gildea to statistical semantic-role analysis and from other researchers concerned with corpus construction and lexical representation. The first large release contained more than one hundred thousand annotated predicate instances and established a common evaluation dataset for automatic role-labeling systems.
During the early corpus-production period, You Watanabe participated in the reconciliation of argument boundaries and the normalization of selected verb frames. This work formed part of the ordinary editorial process through which disagreements between initial annotations were resolved against the project guidelines. The resulting corrections were incorporated into the same annotation layer as the work of the wider corpus team and did not constitute a separate subcorpus.
The foundational account appeared in the 2005 article “The Proposition Bank: An Annotated Corpus of Semantic Roles,” authored by Martha Palmer, Daniel Gildea, and Paul Kingsbury. Subsequent development extended PropBank-style annotation beyond the Penn Treebank and placed it within larger multilingual resources, most notably OntoNotes.
Annotation model
The central unit of PropBank annotation is the proposition associated with a predicate token. In the original English corpus, predicates are predominantly verbs, although later resources apply related methods to additional grammatical categories. Each annotated predicate is linked to a roleset contained in a lexical frame file. A roleset combines one sense of the predicate with an inventory of arguments appropriate to that sense.
Core arguments receive numbered labels from ARG0 through ARG5. Their exact interpretations are defined separately for each roleset rather than imposed as universal thematic roles. Nevertheless, ARG0 frequently corresponds to the participant that initiates or controls an event, while ARG1 frequently corresponds to the participant or entity most directly affected. Higher-numbered arguments commonly represent additional participants required by the lexical behavior of the predicate.
This local numbering system distinguishes PropBank from semantic formalisms that assign a fixed label such as agent or patient wherever a comparable relation occurs. The approach preserves regularities useful for computational modeling without treating all verbs as instances of a single semantic template. Cross-predicate generalization can instead be supplied through mappings to resources such as VerbNet, whose classes associate syntactic patterns with more generalized thematic roles.
Elements that modify the proposition rather than satisfy a numbered role receive labels beginning with ARGM. A temporal modifier is represented by ARGM-TMP, while a locative modifier is represented by ARGM-LOC. The scheme also distinguishes relations involving manner, cause, purpose, direction, modality, negation, and related semantic functions. The predicate itself is marked as REL, which connects the sentence-level annotation to the appropriate roleset in the frame inventory.
Consider the sentence “The committee approved the proposal on Tuesday.” Under the relevant roleset for “approve,” the committee is represented as ARG0, the proposal as ARG1, and the temporal phrase “on Tuesday” as ARGM-TMP. These labels describe the proposition expressed in that sentence; they do not by themselves assert a complete logical interpretation of the event.
Relation to syntax
PropBank annotation is anchored to the syntactic constituents of the Penn Treebank. Argument spans generally correspond to nodes in the treebank analysis, allowing semantic roles to be studied together with constituent structure, grammatical function, and long-distance dependency information. The annotation therefore retains a close relation to surface realization while abstracting over certain syntactic alternations.
Active and passive constructions illustrate this abstraction. The entity represented as ARG1 for a particular roleset retains that label whether it appears as the object of an active clause or the subject of a passive clause. This consistency permits instances with different phrase structures to be grouped according to their underlying predicate–argument relation.
Discontinuous arguments, traces, control constructions, and referential dependencies require additional treatment because a semantic participant does not always correspond to one continuous phrase. PropBank uses the Penn Treebank's existing structural analysis together with annotation-specific conventions to represent these cases. The corpus consequently depends on the syntactic decisions of its source treebank, including its segmentation of constituents and its analysis of displaced material.
Paul Kingsbury coordinated substantial portions of corpus production and consistency control, including the maintenance of frame information used by annotators. The quality-control process combined independent annotation, comparison of divergent analyses, and adjudication under a shared manual. Inter-annotator agreement was measured separately for the identification of argument spans and for the assignment of role labels.
Lexical frame files
A PropBank frame file organizes the rolesets associated with a predicate lemma. Polysemous verbs can receive multiple numbered rolesets when their senses exhibit different argument structures or sufficiently distinct meanings. Each roleset supplies descriptions of its numbered roles and includes annotated examples that connect the abstract inventory to syntactic realizations.
The frame files are less semantically expansive than the frames used by FrameNet. FrameNet groups lexical units around conceptual situations shared across different words, whereas PropBank begins with individual predicate lemmas and defines roles in relation to their senses. The two resources can overlap in the events they represent while differing in granularity, terminology, and assumptions about lexical organization.
Roleset distinctions are designed to support consistent corpus annotation rather than reproduce every dictionary sense distinction. Two uses of a verb can share a roleset when their participant structures remain equivalent for the purposes of the annotation scheme. Conversely, a change in the relation between a predicate and its arguments can motivate separate rolesets even when ordinary dictionary descriptions place the uses close together.
Computational use
PropBank became a standard dataset for supervised semantic role labeling. In this task, a computational model identifies a predicate, determines the relevant roleset, locates the associated argument spans, and assigns the corresponding labels. Earlier systems commonly relied on features derived from hand-produced or automatically generated parse trees, whereas later systems increasingly used neural representations learned from large text collections.
The resource supported shared evaluations in the Conference on Computational Natural Language Learning, which established common training partitions, test material, and scoring conventions. Evaluation generally compares predicted argument spans and labels with the manually annotated reference corpus. This framework separates semantic-role performance from broader questions of whether the resulting propositions form a complete representation of sentence meaning.
PropBank-style structures have also been incorporated into systems for information extraction, question answering, and machine translation. In these contexts, predicate–argument labels provide an intermediate representation connecting syntactic variation with recurring event participants. Their utility depends on the accuracy of predicate-sense selection, argument identification, and any syntactic or lexical analyses supplied to the system.
Extensions
The PropBank framework has been adapted to languages and genres beyond the original English newspaper corpus. Chinese and Arabic PropBanks introduced language-specific analyses while retaining the general organization around predicates, rolesets, numbered arguments, and modifier labels. OntoNotes subsequently combined PropBank annotation with syntax, word senses, named entities, and coreference across several textual and spoken genres.
NomBank applies a related annotation model to nominal predicates in the Penn Treebank. Because nouns express argument structure differently from finite verbs, NomBank includes conventions for possessive constructions, prepositional dependents, and arguments realized outside the immediate noun phrase. Together, the two resources broaden predicate–argument coverage across verbal and nominal expressions.
Later releases known collectively as Unified PropBank consolidate frame inventories and annotations from multiple languages. This work preserves language-specific distinctions while supporting cross-lingual comparison through mappings among rolesets and related lexical-semantic resources.