Skip to content
Back to search results

Semantic Role Labeling (SRL)

Introduced by
Daniel Gildea & Daniel Jurafsky
Year
2002
Type
NLP shallow semantic parsing task
Font size

What does it do?

Semantic role labeling, introduced by Gildea and Jurafsky in 2002, is a natural-language-processing task that assigns semantic roles — who did what to whom, where, when, and how — to the components around a verb (predicate) in a sentence.

It turns plain text into structured predicate-argument representations and is a foundational tool for event extraction.

The intuition behind the method

A sentence describes an action, but a machine does not automatically know who performed it, what was affected, or where it happened. SRL answers those questions by anchoring on the verb and labeling the surrounding phrases with their semantic role — the agent, the patient, the location, the time, the manner. Instead of treating a sentence as a flat string of words, the method reads it as an event with participants, each carrying a clearly named role.

History

Semantic role labeling was established as an automatic task by Gildea and Jurafsky (2002), who showed semantic roles could be labeled automatically using corpus statistics. The field later moved toward neural models; Shi and Lin (2019) demonstrated that simple BERT-based models reach strong performance on SRL and relation extraction, reflecting the broader shift to pretrained language models.

When to use it

Use SRL when you have sentence-level text and need to extract structured who-did-what-to-whom information rather than just a topic or sentiment label. It fits English or another supported language, works best with a PropBank- or FrameNet-style ontology to define the roles, and needs at least a modest corpus to be worthwhile. If your text is in an unsupported language or no role ontology applies, SRL is not the right tool.

Limitations and considerations

  • Quality depends on the language model supporting the text's language — works on English or other supported languages.
  • A suitable role ontology (PropBank or FrameNet) should be chosen up front; without one the role set is undefined.
  • It is a moderately advanced task that operates at sentence level and needs enough data to be meaningful.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Turns free text into explicit predicate-argument structure, exposing who did what to whom.
  • Serves as a foundational building block for event extraction, question answering, and information extraction.
  • Grounds extracted meaning in established ontologies (PropBank, FrameNet) rather than ad-hoc labels.

Applications

  • Event extraction — identifying actions and their participants in news and historical text.
  • Information extraction pipelines that need structured who/what/where/when from sentences.
  • Supporting question answering by exposing the predicate-argument structure of a passage.
  • Analysing large text corpora in social science, humanities, and health research where event structure matters.

How it works

How it works

1

The sentence is scanned to find the predicates — typically the verbs that anchor an event. Each predicate becomes the centre around which roles will be organised, because the meaning of a role is defined relative to the action it serves.

2

For each identified predicate, the spans of text that act as its arguments are located. This step decides which phrases participate in the event before deciding what part they play, separating relevant constituents from incidental words.

3

Each argument span is assigned a semantic role label — agent, patient, location, time, manner, and so on — typically following the PropBank or FrameNet ontology. This is where the who/what/where/when/how questions are answered for every participant.

4

The labeled predicates and arguments are assembled into a predicate-argument structure for the sentence. This structured event representation is the output that downstream tasks such as event detection and question answering build on.

Common pitfalls

  • Applying an English-trained model to text in an unsupported language degrades labeling badly.
  • Skipping the choice of a PropBank or FrameNet ontology leaves role definitions inconsistent across sentences.
  • Errors in predicate identification cascade — if the verb is missed, none of its arguments get labeled.
  • Running SRL on too little text (under ~20 sentences) gives unstable, unrepresentative output.

Frequently asked questions

How is SRL different from named-entity recognition?

NER identifies and types entities (people, places, organisations) regardless of their role in an action. SRL goes further: it anchors on the verb and labels which entity is the agent, which is the patient, and where, when, and how the action happened — capturing the event structure rather than just the entities.

What are PropBank and FrameNet?

They are the two main ontologies that define the inventory of semantic roles SRL assigns. Choosing one up front gives consistent, comparable role labels across all sentences; without an ontology the role set is undefined.

Does SRL work in languages other than English?

It works on English or another supported language, provided a model exists for that language. Applying an English-trained model to an unsupported language is a common cause of poor results.

Why is SRL described as a foundational tool for event extraction?

Because an event is essentially a predicate with participants. By producing predicate-argument structure, SRL hands downstream event-detection and information-extraction steps exactly the who/what/where/when they need.

Related methods

Sources

  1. Gildea, D. & Jurafsky, D. (2002). Automatic Labeling of Semantic Roles. Computational Linguistics, 28(3), 245-288.

    Open DOI reference
    Copied.
  2. Shi, P. & Lin, J. (2019). Simple BERT Models for Relation Extraction and Semantic Role Labeling. arXiv:1904.05255.

    Open publication reference
    Copied.