Is disambiguation the same as translation?

No. Translation may require disambiguation, but it also generates an expression in another language.

Can more context make the answer certain?

Sometimes it resolves the ambiguity, but misleading or incomplete context can still leave uncertainty.

Is a dictionary sufficient?

It supplies candidate senses, while choosing among them requires evidence about the particular use.

AI Glossary / Graphs and symbolic AI / Advanced

Disambiguation

Definition

Disambiguation is the process of determining which interpretation of an expression is intended when more than one is possible. In natural language understanding, it uses context and relevant knowledge to distinguish candidate meanings. The available evidence may support one interpretation strongly, weakly, or not enough to choose responsibly.

What it means

Ambiguity occurs at several levels. A word can have different senses, a sentence can have different grammatical interpretations, and a name can identify different people or organizations. Disambiguation concerns the choice among these possibilities, rather than simply recognizing that the words are present.
For example, “Please check the draft” might concern a document, an air current, or a selection process. A nearby reference to tracked changes supports the document interpretation. Without that context, choosing the most common sense is a useful baseline but does not establish the speaker's intention.

How it works

A system identifies candidate interpretations and compares them with contextual evidence. For word-sense disambiguation, candidates may come from a sense inventory. For named-entity disambiguation, candidates are entities in a knowledge base. Candidate generation matters: ranking cannot select the correct entity if it was never offered.
Useful context may include nearby words, grammatical relationships, earlier conversation, and relevant domain information. A model or rule system scores the alternatives. The application then decides whether to select one, retain several, or request a missing detail.
These tasks have related but different outputs. Resolving a name to a database identifier is not the same as selecting a dictionary sense. Coreference resolution identifies expressions that refer to the same entity, including references within a document. A practical language system may need several of these operations together.

Three practical examples

Internal search: An employee searches for “the launch deck.” Context about a presentation project favors slides rather than a ship's deck. The search system can use that interpretation to retrieve relevant files while allowing the employee to change the scope.
Record linking: A report mentions “Alex Morgan,” a name shared by several contacts. An associated employer or project can distinguish candidates. Automatically merging records solely because their displayed names match can combine information about different people.
Instruction understanding: “Send Jordan the revised version” requires identifying Jordan and the relevant version. A conversation may supply both references, but if two projects are active, the assistant should obtain the missing distinction before treating one file as the intended object.

Think of it this way

Think of reading a sign through a window that reveals only part of the scene. The letters may be clear while their meaning remains uncertain. More surrounding detail can settle the interpretation; confidence based only on familiarity can still point to the wrong scene.

Worked example

Create an evaluation set of 100 messages containing the word “draft.” Reviewers identify 70 as referring to a document and 30 as referring to an air current. Assume these senses cover this simplified dataset and its reference labels are reliable.
A baseline always chooses the document sense. It correctly labels 70 messages, giving accuracy 70/100 = 70%. It makes no correct decisions on the 30 air-current cases, despite its apparently reasonable overall score.
A context-based system correctly labels 63 of the 70 document cases and 24 of the 30 air-current cases. Its overall accuracy is (63 + 24)/100 = 87%. The document-sense accuracy is 90%, while the air-current-sense accuracy is 80%. These separate results show where errors remain.
Now apply an illustrative abstention rule to that system's decisions. Suppose it answers 80 messages and is correct on 75 of those, leaving 20 for clarification or review. Its answered-case accuracy is 75/80 = 93.75%, and its coverage is 80/100 = 80%. It has 5 errors among answered cases. The original 87 correct decisions can be reconciled as 75 retained plus 12 among the 20 deferred cases.
The higher answered-case accuracy does not mean it correctly resolves 93.75 of every 100 incoming messages. It directly resolves 75 correctly, answers 5 incorrectly, and defers 20. Whether that tradeoff is useful depends on the consequences of errors and the cost and effectiveness of resolving deferred messages.

Illustrative case study

Fictional Cedar Support processes requests mentioning “the terminal.” Its older rule always routes them to an IT queue. Reviewers find that some messages instead concern a transport terminal in the company's facilities operations.
The team builds a labeled sample covering both meanings, including short messages with little context. It evaluates a context-aware classifier against the existing default and introduces a clarification path for unresolved cases. Reviewers also check whether the company's internal vocabulary has meanings missing from its initial labels.
The revised workflow reports correctness by meaning and the proportion of requests deferred. Its success measure is improved routing of real requests, rather than a confident label attached to every ambiguous word.

Common misconceptions and mistakes

“The most frequent meaning is always intended.” Frequency supplies a prior preference, not certainty.
“Recognizing a name identifies the person.” Several entities can share a name.
“A high model score proves the interpretation is correct.” Scores need appropriate evaluation and may not be calibrated probabilities.
“Every expression has one recoverable meaning.” The context may be insufficient, or ambiguity may be intentional.
“Abstaining improves every metric.” It can improve answered-case accuracy while reducing coverage and increasing review work.

Business use

Define what a correct interpretation enables: retrieving a document, selecting a record, routing a request, or preparing an action. Test the system on the organization's actual vocabulary and on minority meanings, not only the most frequent sense.
Keep unresolved cases visible. Report coverage alongside accuracy when the system can defer, and evaluate the clarification process itself. For actions with material consequences, verify the relevant identity or object before execution rather than treating an interpretation score as confirmation of intent.

Check your understanding

Why does the baseline reach 70% accuracy?

Answer: It always selects the majority sense.

What does 80% coverage mean?

Answer: The system answers 80 of the 100 cases directly.

Why is 93.75% not end-to-end resolution accuracy?

Answer: It is calculated only over the answered subset.

Related terms

Learn next

Collect several genuinely ambiguous workplace expressions. List plausible interpretations and identify the smallest contextual detail that would distinguish them. Evaluate both wrong interpretations and unresolved cases when comparing systems.

Sources

AIMA glossary, disambiguation: https://github.com/aimacode/aima-glossary/blob/master/glossary.md Chang, Spitkovsky, Manning, and Agirre, A comparison of Named-Entity Disambiguation and Word Sense Disambiguation: https://nlp.stanford.edu/pubs/chang2016entity.pdf Stanford NLP Group, coreference resolution overview: https://nlp.stanford.edu/projects/coref.shtml

Previous: Dirty data Next: Discordant pairs