Can AI perform subject indexing?

Yes, but its assignments should be evaluated for the collection and policy.

Is a subject heading the same as a classification number?

Both can express subject information, but they serve different representational and organizational roles.

Does controlled vocabulary eliminate judgment?

No. People or models must still decide which concepts appropriately describe a resource.

AI Glossary / Retrieval and knowledge systems / Intermediate

Subject indexing

Definition

Subject indexing is the process of describing what an information resource is about through assigned terms, headings, codes, or other subject indicators. It helps people retrieve and organize resources by topic, including when authors use different words for similar concepts. Indexing may be performed by people, automated systems, or a combination, and can use controlled vocabularies or less controlled terms.

What it means

A document may discuss a concept without using the exact phrase a searcher enters. Subject indexing creates an additional description that connects the resource to that concept. It is therefore more than simply counting the words printed in the document.
The description reflects an indexing policy: which subjects deserve representation, how specifically they should be named, and how much of the resource should be covered. These decisions influence what users can find. A brief mention and a central theme need not receive the same treatment.

How it works

The first task is conceptual analysis: examine the resource and identify its important subjects. Titles, abstracts, headings, and other evidence can help, but a misleading title should not substitute for examining relevant content.
The next task translates those subjects into the chosen indexing language. A controlled vocabulary can connect alternate expressions to an authorized concept and provide broader, narrower, or related relationships. Free keywords permit flexibility but may create inconsistent labels for the same subject.
Specificity concerns how precisely a selected term represents a subject. Exhaustivity concerns how many aspects are represented. Assigning many broad terms is not necessarily either specific or useful. The appropriate balance depends on the collection, users, and policy.
Automated systems can propose or assign labels from text, but their outputs require evaluation against intended subject coverage. Maintain vocabulary versions and records of assignment decisions. A subject label can be stored separately from the full-text search index, and the search interface may combine both forms of evidence. Hierarchical expansion, if offered, must follow the vocabulary and query rules actually implemented.

Three practical examples

A research library assigns controlled subject headings so readers can discover publications across varying author terminology.
A company knowledge base labels procedures by business process, equipment type, and topic under a documented policy.
A digital archive uses AI-assisted suggestions that editors review for centrality, specificity, and consistency before accepting them.

Think of it this way

Think of a museum catalog describing what each object concerns, rather than listing only the words written on its label. Visitors can then explore a topic even when the labels use unfamiliar expressions. The descriptions help navigation, but they still depend on careful judgment about the objects.

Worked example

Consider an original, invented evaluation collection of eight documents. A subject expert identifies documents 1, 2, 3, and 4 as relevant to the topic battery recycling. This four-document set is the reference for this exercise.
An exact keyword search retrieves documents 1, 2, and 5. Two of the three retrieved documents are relevant, so precision is 2/3, approximately 66.67%. It finds two of the four relevant documents, so recall is 2/4 = 50%.
A subject-heading search retrieves documents 1, 2, 3, 4, and 6. Four of five retrieved documents are relevant, giving precision of 80%. It retrieves all four relevant documents, giving recall of 100%.
Now inspect the assignments. Document 6 only mentions recycling in passing, so under the stated indexing policy it should not have received this topic label. Removing that inappropriate assignment makes the subject search return exactly documents 1 through 4, giving 100% precision and recall on this toy collection.
The arithmetic illustrates how consistent labels can recover relevant resources that keyword matching misses, while inappropriate assignments can add false matches. It does not prove that subject indexing always beats full-text search. The result depends on the invented collection, the query, the expert reference, and the exact retrieval rules.
If the policy instead intended to capture every brief mention, the judgment about document 6 might change. Evaluation must therefore state what counts as relevant rather than treating the indexing policy as invisible.

Illustrative case study

Fictional case: Spruce Learning organizes internal training materials. Authors use different expressions for the same process, so employees miss useful modules when searching with their own department's terminology.
Editors create a small controlled vocabulary and review the main subjects of each module. An automated model suggests candidate labels, but reviewers check whether each topic is central enough under the agreed policy. The team tests known employee questions and examines both missing results and irrelevant matches.
This hypothetical case illustrates assisted indexing and policy review. It reports no real deployment or measured improvement.

Common misconceptions and mistakes

A subject label is not necessarily a word that appears in the source. It can represent a concept expressed through other language.
Subject indexing is not the same as building an inverted file for every token. The former describes aboutness; the latter organizes searchable occurrences, although both can support retrieval.
More labels do not automatically improve discovery. Excessively broad or incidental assignments can produce noisy results. Controlled terms also require maintenance and can reflect historical or cultural assumptions, so consistency alone does not establish neutrality or adequacy.

Business use

Subject indexing can help businesses organize growing libraries of policies, research, training, and product knowledge. It is most valuable when users need stable topic access across inconsistent wording or when a curated structure supports browsing.
Begin with representative search tasks and clear assignment rules. Track coverage, inappropriate labels, and reviewer agreement rather than only throughput. Plan how revised terms and changed content will be reindexed. Retain full-text retrieval where it complements subject access, especially for unexpected names, quotations, or details outside the indexing policy.

Check your understanding

Why did recall rise in the example?

Answer: Subject access retrieved relevant documents 3 and 4 that the exact keyword search missed.

Why was document 6 a problem?

Answer: Its incidental mention did not meet the stated subject-assignment policy.

What distinguishes specificity from exhaustivity?

Answer: Specificity concerns precision of description; exhaustivity concerns the extent of subject coverage.

Related terms

Learn next

Learn how controlled vocabularies represent concepts and relationships. Then practice assigning subjects under an explicit policy and compare the resulting retrieval behavior with keyword search on a small, reviewed collection.

Sources

Primary professional guidance: Library of Congress, Subject Headings Training, Module 1.4, How Do We Determine Aboutness?: https://www.loc.gov/catworkshop/lcsh/PDF%20scripts/1-4-Aboutness.pdf Primary vocabulary resource: National Library of Medicine, Medical Subject Headings: https://www.nlm.nih.gov/mesh/meshhome.html

Previous: Subgroup analysis Next: Subject-verb agreement