AI Glossary / Natural language and speech / Advanced

Semantic role labeling

Definition

Semantic role labeling, or SRL, is the task of identifying a predicate's arguments and assigning labels that describe their roles in the event or situation. It converts parts of a sentence into a structured account of participation, such as who sent something, what was sent, to whom, and when, under a specified annotation scheme.

What it means

Recognizing Nora and Ivo as people is different from identifying which person sent a parcel and which received it. SRL connects expressions to a particular predicate through their semantic roles.
The task must be specified carefully. Some evaluations supply the predicate and ask only for its arguments and labels. A complete pipeline may also need to find predicates and disambiguate their senses. Span-based SRL identifies full argument phrases; dependency-based SRL identifies their heads and roles. Scores from these settings are not interchangeable. A useful system description states what information was supplied and what the model had to infer.

How it works

A typical supervised approach learns from sentences annotated with a declared resource such as PropBank or FrameNet. The system represents the sentence together with the target predicate, identifies candidate arguments, and predicts their roles. Those decisions may be separate stages or jointly modeled.
One span-based approach uses BIO tags: B starts an argument, I continues it, and O marks tokens outside the argument spans for that predicate. The role name accompanies B and I. Contextual encoders can provide representations for prediction; an external syntactic parser is useful in some designs but is not mandatory for every SRL model.
Evaluation must check both the argument boundaries and the role labels, using the conventions of the chosen task.

Three practical examples

• A document search tool answers who delivered a shipment by finding the participant assigned the relevant role, rather than returning any person mentioned nearby.
• A summarization check compares participant assignments in a source sentence and its summary. This can expose a swapped sender and recipient despite fluent wording.
• A workflow analyst organizes descriptions of requests, approvals, and transfers. Each predicate receives its own analysis so that participants from separate actions are not silently combined.

Think of it this way

Think of completing an event form from a sentence. Finding a phrase determines what text belongs in a box; labeling its role determines which box it belongs in. Copying only part of the address is a boundary error. Putting a complete address in the date box is a role error. Both matter even when many individual words were handled correctly.

Worked example

Take eight whitespace-separated tokens: Nora(1) sent(2) the(3) red(4) parcel(5) to(6) Ivo(7) yesterday(8). Supply sent as the target predicate. For this original toy scheme, define four gold arguments: Sender at 1-1, Item at 3-5, Recipient at 6-7, and Time at 8-8. Including to in Recipient is an explicit convention of this exercise, not a claim about every annotation resource.
Suppose the model predicts Sender at 1-1, Item at 3-4, Recipient at 6-7, and Location at 8-8. Only Sender and Recipient match both boundaries and labels. Exact labeled precision is 2/4 = 50%; recall is also 2/4 = 50%; F1 is 50%.
Ignoring role labels, three of four spans match, giving 75% unlabeled precision and recall. Under BIO encoding with the supplied predicate tagged O, six of eight token tags match: 75% token accuracy. The missing parcel token and the wrong role on yesterday cause the two mismatches. Thus a fairly high token score does not imply that the complete arguments are correct. All figures describe this invented example, not measured model performance.

Illustrative case study

Fictional platform DispatchNote uses SRL to organize delivery messages. Its demonstration succeeds on short active sentences, but a review set includes passive wording, omitted participants, and messages describing two separate transfers.
Reviewers record whether each failure comes from a wrong predicate, an incomplete phrase, a wrong role, or a later reference-resolution step. They also separate supplied-predicate tests from full-pipeline tests. This reveals which component needs improvement instead of attributing every problem to one undifferentiated language-understanding score.
The fictional team keeps source text beside extracted fields and marks unresolved information explicitly. The scenario illustrates evaluation design without claiming a production result.

Common misconceptions and mistakes

SRL is not named-entity recognition: knowing an expression is a person or organization does not determine its role. It is also not complete semantic understanding. Reference, event identity, negation, modality, and factual verification may require additional treatment.
Do not interpret numbered argument labels without their predicate-sense definitions. Avoid treating every grammatical subject as the actor. Passive and other constructions break that shortcut.
Do not report token accuracy as if it were exact argument F1, or compare results with different predicate assumptions. When a phrase boundary is wrong, getting most of its words right may still fail the required exact-match criterion.

Business use

SRL can provide useful structure for search, document analysis, and consistency checking when the roles fit the application. Evaluate on representative writing styles and on the specific participant reversals users need to avoid. Preserve predicate links and source spans in the output. If structured fields drive later decisions, assess that downstream task as well: a better role score does not automatically establish better business outcomes.

People also ask

Does SRL always need a parse tree?

No. Some methods use syntax explicitly, while others predict roles from contextual representations without an external parser.

Can a sentence have several analyses?

Yes. Each predicate can have its own arguments, and some expressions participate in more than one structure.

Does SRL verify that an event happened?

No. It analyzes linguistic structure; a hypothetical or denied event still has participants.

Check your understanding

Why is the predicted Item incorrect in the exercise?

Answer: Its span omits parcel and therefore fails the exact boundary requirement.

Why does yesterday pass the unlabeled span check but fail the labeled check?

Answer: Its boundaries match, but Location differs from the gold Time role.

What extra responsibility arises when predicates are not supplied?

Answer: The system must also identify which expressions are predicates, with the relevant sense decisions required by its task.

Related terms

• Semantic role
• Predicate (logic)
• BIO tagging
• predicate
• argument identification
• argument classification
• span extraction
• PropBank
• FrameNet
• dependency parsing

Learn next

Annotate a small set with declared span conventions and predicate assumptions. Compare token accuracy, unlabeled span scores, and labeled argument F1. Then add passive constructions and multiple predicates to see which errors the simpler test concealed.

More in Natural language and speech

• Semantic interpretation - Semantic interpretation is the process of turning a natural-language utterance into a representation of its meaning that a computer system can use.
• Semantic neural network - A semantic neural network is a research-specific network model that connects representations of meaning with neuron-like processing.
• Semantic parsing - Semantic parsing is the task of converting a natural-language utterance into a formal representation of its meaning.
• Semantic similarity - Semantic similarity is the degree to which two words, phrases, sentences, or other representations resemble one another in meaning.
• Semantic space - A semantic space is a representation in which words, texts, or concepts occupy positions intended to reflect aspects of meaning.
• Semantometrics - Semantometrics is an approach to research evaluation that uses the semantic content of scholarly publications, often together with their relationships in...

Sources

Jurafsky and Martin, Speech and Language Processing, Chapter 22, section 22.6, SRL methods and evaluation: https://web.stanford.edu/~jurafsky/slp3/22.pdf Shi and Lin, Simple BERT Models for Relation Extraction and Semantic Role Labeling, 2019, section 3: https://arxiv.org/html/1904.05255v1 The parcel annotation, scores, and DispatchNote scenario are original illustrations. Review note: Chapter 22 was inaccessible; Shi and Lin’s primary paper was consulted.

Previous: Semantic role Next: Semantic service-oriented architecture