Semantic role labeling (SRL), also known as shallow semantic parsing, is the task of identifying the predicates in a sentence and labeling their arguments with semantic roles such as Agent, Patient, Theme, Instrument, Location, and Time. SRL provides a level of semantic representation intermediate between syntactic parsing and full logical form, capturing "who did what to whom" without requiring a complete formal semantic analysis. The task draws on Fillmore's case grammar and frame semantics, and is evaluated using resources like PropBank and FrameNet.
Role Inventories and Resources
PropBank-style:
Predicate: chased
Arg0 (Agent): The cat
Arg1 (Patient): the mouse
ArgM-DIR (Direction): into the garden
FrameNet-style:
Frame: Cotheme (pursuing variant)
Theme: The cat
Cotheme: the mouse
Goal: into the garden
PropBank defines roles on a verb-by-verb basis using numbered arguments (Arg0 through Arg5) plus modifier roles (ArgM-TMP for temporal, ArgM-LOC for locative, etc.). Arg0 typically corresponds to the agent or proto-agent, and Arg1 to the patient or theme. FrameNet uses semantically specific role labels organized by frame. NomBank extends PropBank-style annotation to nominal predicates. These resources provide the training data for supervised SRL systems and define the evaluation standards for the CoNLL shared tasks on SRL.
Models and Architectures
Early SRL systems used pipeline architectures with separate stages for predicate identification, argument identification, and argument classification, relying on syntactic parse features. The CoNLL 2004/2005 shared tasks established SRL as a benchmark, with systems using SVMs and maximum entropy classifiers over syntactic features. Modern neural SRL systems use deep BiLSTM or Transformer encoders, often achieving strong results without explicit syntactic features. He et al. (2017) showed that a deep BiLSTM with highway connections and constrained decoding achieves state-of-the-art performance, and BERT-based models have further advanced the field.
SRL comes in two flavors: span-based (identifying argument spans as contiguous text sequences) and dependency-based (identifying argument heads). Span-based SRL, used in the CoNLL 2005 shared task, labels constituent spans. Dependency-based SRL, used in CoNLL 2008/2009, labels syntactic heads. The span-based formulation is more natural for end-to-end neural models, while dependency-based SRL aligns better with dependency parsing and cross-lingual transfer. End-to-end span-based models that jointly predict predicates and arguments have become the dominant approach.
Applications
SRL output is used as a feature or intermediate representation in many NLP tasks. In question answering, semantic roles help match questions to answer spans: a "who" question seeks an Agent, a "what" question seeks a Patient or Theme, and a "where" question seeks a Location. In information extraction, SRL identifies event participants in a domain-independent way. In machine translation, preserving semantic roles across languages improves translation quality, particularly for argument reordering. Textual entailment systems use SRL to compare the predicate-argument structures of premise and hypothesis.
Cross-lingual SRL transfers role labeling knowledge from resource-rich to resource-poor languages using multilingual word embeddings or annotation projection. Universal PropBank and the CoNLL 2009 multilingual SRL dataset provide resources for evaluating cross-lingual transfer. Recent work on SRL has focused on implicit arguments (roles that are not syntactically realized but are understood from context), abstract meaning representations that integrate SRL with coreference and discourse structure, and multi-sentence SRL that tracks role fillers across document contexts.