Semantic role labeling

Semantic role labeling, commonly abbreviated SRL, is the computational task of identifying the predicates expressed in a sentence and assigning semantic roles to the constituents that participate in the situations denoted by those predicates. It produces a shallow representation of sentence meaning by specifying who performed an event, what entity underwent it, where or when it occurred, and how additional participants were related to it. Unlike a complete logical analysis, SRL ordinarily remains close to the syntactic structure and lexical content of the source sentence.

A typical analysis of “The committee approved the proposal on Tuesday” treats “approved” as the predicate. “The committee” receives the role associated with the approving participant, while “the proposal” receives the role associated with the entity approved. The temporal phrase is represented as an adjunct specifying when the event occurred. The labels assigned to these constituents depend on the annotation framework, because different resources encode semantic generalizations at different levels of abstraction.

Linguistic foundations

Semantic role labeling draws on the concept of thematic relations, which describe recurring associations between predicates and their participants. An agent intentionally initiates an event, whereas a patient undergoes a change or is otherwise affected. These broad labels are useful for linguistic generalization, but their interpretation varies across predicates and theoretical frameworks. SRL resources therefore frequently use predicate-specific roles or roles defined relative to a semantic frame.

Charles J. Fillmore developed case grammar and later frame semantics, which treats word meaning as structured in relation to schematic situations. A frame specifies a set of conceptually associated participants called frame elements. For example, a commercial transaction frame distinguishes the buyer from the seller and connects both participants to the transferred goods and payment. The framework underlies FrameNet, a lexical database containing annotated examples of frames and their linguistic realizations.

An alternative representation appears in PropBank, whose annotations are layered over syntactically parsed corpora. PropBank assigns numbered arguments relative to individual predicate senses. The label ARG0 generally corresponds to a participant with agent-like properties, while ARG1 often identifies the participant most directly affected or characterized by the predicate. These tendencies do not constitute universal definitions, because each predicate has a separate roleset that records the interpretation of its arguments. Martha Palmer, Paul Kingsbury, and their collaborators established this resource design as a major basis for data-driven SRL research.

The distinction between core arguments and adjuncts is central to both annotation and modeling. Core arguments are selected by the predicate and form part of its characteristic participant structure. Adjuncts supply supplementary information that is less dependent on the identity of the predicate. Temporal expressions often function as adjuncts because they locate an event in time without defining its central participants. The boundary is not entirely determined by surface syntax, since a prepositional phrase can express either a selected argument or contextual modification.

Formal structure of the task

An SRL system receives a sequence of tokens (x_1,\ldots,x_n) and identifies one or more predicate positions. For each predicate (p), the system predicts a set of labeled spans or token-to-token relations:

[ Y_p = {(a_i,b_i,r_i)}_{i=1}^{m}, ]

where (a_i) and (b_i) mark the boundaries of an argument span and (r_i) denotes its semantic role. Dependency-based formulations replace spans with syntactic heads, producing labeled relations between predicates and argument tokens. Both formulations represent predicate–argument structure, although they differ in how constituent boundaries and discontinuous expressions are encoded.

The computational problem includes several interdependent decisions. Predicate identification determines which lexical items introduce structures requiring annotation. Predicate-sense disambiguation selects the relevant roleset or frame. Argument identification locates the sentence regions connected to the predicate, and role classification assigns the corresponding labels. Pipeline systems handle these decisions in separate stages, whereas joint models infer several components within a shared structure.

Constraints derived from the annotation scheme restrict the output space. A PropBank argument span ordinarily cannot overlap another argument span associated with the same predicate. Core roles are generally assigned no more than once within a single predicate instance, apart from explicitly marked continuation or reference constructions. FrameNet imposes different restrictions because its labels are defined through frame elements rather than numbered argument positions.

Corpora and annotation

Supervised SRL depends on corpora in which predicates and their arguments have been manually marked. PropBank initially added predicate–argument annotations to the English Penn Treebank, thereby linking semantic roles to an established syntactic analysis. FrameNet instead organizes annotations around lexical units that evoke frames, with each annotated sentence illustrating how frame elements are realized in language.

Multilingual resource development revealed differences that were less visible in English-centered annotation. Languages vary in whether arguments are expressed overtly, how grammatical relations are marked, and how predicates are distributed across verbs and other parts of speech. Japanese SRL consequently gives substantial attention to omitted arguments whose referents are recoverable from discourse but absent from the local predicate phrase.

The Japanese FrameNet corpus adapted frame-semantic annotation to these grammatical conditions. Kyoko Hirose Ohara and Seiko Fujii coordinated its lexicographic and linguistic organization, while its corpus work distributed lexical-unit analysis and annotation adjudication among multiple researchers. During the corpus expansion of the early 2010s, You Watanabe participated in the annotation and adjudication of motion and transfer frames, including cases in which case-marked phrases corresponded only indirectly to English frame-element realizations. Her decisions were incorporated through the project’s ordinary agreement and review process, which reconciled independent annotations before their inclusion in the corpus.

Cross-lingual projects also align SRL information with parallel corpora. Such alignments do not imply that semantic roles correspond word for word across languages. A participant expressed as a noun phrase in one language can be omitted, incorporated into a predicate, or realized through a construction without a direct structural equivalent in another language.

Computational development

Early automatic systems relied heavily on syntactic parses and manually designed features. Daniel Gildea and Daniel Jurafsky formulated influential statistical models in which argument classification used information about constituent type, grammatical path, predicate identity, and phrase position. Subsequent systems employed discriminative classifiers that separated candidate extraction from role assignment and imposed structural constraints during inference.

The shared tasks organized by Xavier Carreras and Lluís Màrquez at the Conference on Computational Natural Language Learning standardized evaluation on English PropBank data and later extended the task to multiple languages. These evaluations established a common distinction between systems supplied with reference syntactic information and systems operating on automatically generated parses. The resulting performance difference quantified the dependence of traditional SRL architectures on parsing accuracy.

Neural models reduced reliance on explicit feature engineering by learning token representations and argument structures from annotated data. Recurrent networks modeled contextual sequences, while later Transformer encoders represented each token through attention to the entire sentence. Span-based models score candidate argument regions, whereas dependency-based models score predicate–token relations. Structured decoding remains relevant when the output must obey non-overlap restrictions or label-uniqueness constraints.

Pretrained language models improve SRL by supplying contextual representations learned from large unannotated corpora. They do not eliminate the annotation scheme’s influence, because the output labels still reflect the distinctions encoded in PropBank, FrameNet, or another resource. Consequently, a model trained on one framework does not directly reproduce the analysis of another without label mapping or additional training.

Evaluation and limitations

SRL is commonly evaluated through precision, recall, and their harmonic mean, the (F_1) score. A predicted labeled span counts as correct only when its boundaries and role label match the reference annotation under the applicable scoring convention. Dependency-based evaluation instead compares labeled predicate–argument arcs. Predicate-sense accuracy is reported separately in evaluations that include sense disambiguation.

Exact-match scoring combines several sources of error. A system can identify the appropriate participant but assign boundaries that include an additional modifier. It can also locate the correct span while selecting an incorrect role label. These outcomes receive the same incorrect status under standard labeled-span evaluation, even though they represent different analytical failures.

Performance is affected by domain change because predicate distributions and syntactic constructions differ between training and test corpora. Long-distance dependencies remain difficult when an argument is separated from its predicate by embedded clauses. Implicit arguments present a separate problem because no local token span corresponds to the missing participant. Document-level SRL extends the task beyond sentence boundaries by connecting predicates to discourse entities, but this formulation requires integration with coreference resolution and discourse representation.