What is information science?

What the field studies

Information science covers six concerns. Every serious AI system has to answer each of them, whether or not anyone on the team knows the field exists.

Organizing knowledge. A domain has to be divided into things and relationships before anyone can move through it. Classification schemes, taxonomies, thesauri, and ontologies all live here. The field’s oldest rule on the subject comes from Charles Cutter’s cataloguing rules, first published in 1876: the convenience of the public comes before the ease of the cataloguer. A good structure is built from the questions people will ask. The shape of the collection is secondary.

Describing it. Metadata is the data about a document that lets a system handle it without reading it: who made it, when, about what, in what form, for whom. Cataloguing rules and authority files exist because free-text description stops working once a collection outgrows the people who remember it. Dublin Core, named for the 1995 workshop in Dublin, Ohio where it began, settled on a 15-element core for describing anything on the web. A chunk in a vector index with no source, date, or type is an uncatalogued document, and it fails the same way.

Finding it. Information retrieval is the sub-field that studies matching a need to a document. Calvin Mooers coined the term in 1950. Boolean search, term weighting, the vector space model, and relevance feedback all come from here. Embedding search descends directly from Gerard Salton’s work on representing documents as vectors, published in full in 1975.

Judging whether finding worked. Precision and recall predate machine learning by decades. They became measurable in the Cranfield experiments at the College of Aeronautics in England, run by Cyril Cleverdon between 1957 and 1966, which built the first test collections: a fixed set of documents, a fixed set of queries, and human relevance judgments. The Text REtrieval Conference (TREC), running since 1992, carried that method into the web era. Anyone building an evaluation set for a retrieval system is running a Cranfield experiment.

How people seek and use information. Information behavior studies what people do when they need to know something: how they phrase the need, when they stop searching, what they trust, and how the need changes as they learn. Robert Taylor’s 1968 paper on question negotiation described four stages a question passes through before it reaches a librarian. The last stage, the compromised question, is the one the system receives. That is why a query log looks nothing like a taxonomy, and why an agent given a vague goal behaves differently from one given a well-formed question.

Knowing where it came from. Provenance and authority cover the chain from a claim back to its origin and whether that origin deserves trust. Eugene Garfield’s Science Citation Index, first published in 1964, turned the citation into a traceable link, and authority control gives each author and organization exactly one name. In an AI system this is the difference between an answer and a verifiable answer.

A short history

The field grew out of two problems: too many documents, and documents that could not be found.

In 1895 the Belgian lawyer Paul Otlet and Henri La Fontaine set out to index all of recorded knowledge on cards. The project produced the Universal Decimal Classification and the Mundaneum, which by the 1930s held millions of index cards and answered written queries by post. Otlet’s Traité de documentation (1934) described a desk with a screen that retrieved documents on demand from a remote store. In India, S. R. Ranganathan published the Colon Classification in 1933 and with it facet analysis: describing a subject by combining independent dimensions instead of placing it in one branch of a tree. Faceted filters on e-commerce sites descend from that idea.

Vannevar Bush’s 1945 essay “As We May Think” imagined the memex, a desk that stored a personal library on microfilm and let its owner build and share associative trails between documents. Hypertext and the web cite it as an ancestor. Claude Shannon’s information theory arrived in 1948 and lent the word “information” its prestige. The field took the name a decade later; the Institute of Information Scientists was founded in Britain in 1958. Shannon’s “information” is a measure of signal and says nothing about meaning, and the two uses of the word have been confused ever since.

The 1950s and 60s turned the field quantitative. Mooers named information retrieval. Cleverdon’s Cranfield tests showed that simple keyword indexing performed about as well as elaborate controlled vocabularies, a result the library profession took badly, and the experimental method survived the argument. Salton’s SMART system, begun at Harvard in the early 1960s and moved to Cornell in 1965, introduced the vector representation of documents and relevance feedback (Rocchio, 1965). The American Documentation Institute renamed itself the American Society for Information Science in 1968. The field had stopped being about documents.

From the 1990s the web became the field’s laboratory. TREC standardized large-scale evaluation. Dublin Core gave the web a minimal metadata vocabulary. Tim Berners-Lee’s Semantic Web article in Scientific American (2001) and the standards around it, RDF, OWL, and SPARQL, made ontologies machine-readable at web scale. Google’s Knowledge Graph, launched in 2012, is the industrial form of that work.

Large language models are the latest reader to arrive. They made the field’s problems urgent, because a model will answer confidently from whatever it is given, organized or not.

What it is not

Three neighbours get mistaken for it.

Computer science studies computation: algorithms and the machines that run them. Information science uses computation and studies what the computation handles, the documents and the people who need them. A search engine’s ranking algorithm is computer science. Deciding what “relevant” means for a given user, and measuring whether the ranking achieves it, is information science.

Data science finds patterns in data, mostly numeric, mostly for prediction. Information science organizes and retrieves knowledge, mostly in documents, mostly for a person or agent with a question. The two meet at the data model, and both fail when the definitions underneath it are inconsistent. A feature store with ambiguous column names and a vector index with ambiguous entity names are the same bug.

Library science is the closest neighbour and the one information science grew from. Cleverdon was the librarian of the College of Aeronautics; the experiments that founded retrieval evaluation came out of a library, and the two fields are still often taught together. Library science is the practice of running collections and serving their users. Information science is the study of information in any setting, including ones with no library and no librarian, such as a company’s document store feeding a retrieval system.

Why it matters for AI

A language model is a reader. It reads whatever is placed in its context and produces text in response. Everything that determines whether it answers well happens before the model sees a token: how the material was divided, how it was described, what was retrieved for this question, whether that retrieval was ever measured, how the question was phrased in the first place, and whether the result can be traced to a source. Those are the six concerns above, in order.

So the common failures of AI systems look familiar to anyone trained in the field. Retrieval that returns passages about the right words and the wrong thing is a relevance problem; Cranfield was measuring it in the 1950s. An index that cannot filter by date or document type is a metadata problem cataloguers solved. An agent that cannot tell a customer from an account from a contact has a knowledge organization problem, and ontologists have methods for it. A system that cannot say where an answer came from has a provenance problem, and citation practice exists to address it.

The practical consequence is that the hard part of an AI system is usually not the model. It is the information architecture around the model, and that architecture has a discipline behind it with a century of results. Teams that know the discipline build the knowledge layer deliberately. Teams that do not build it by accident, and then cannot fix it because it was never written down.

From the field to the stack

Each of the six concerns maps onto a layer of the systems this site covers. The topic pages take each one into current tools and practice.

Common questions

Is information science a science? Yes, in the sense that computer science is one. It has an experimental method, the test collection, and a body of results that replicate; the Cranfield finding on keyword indexing has held for sixty years. It is also a design discipline, because organizing knowledge means making choices.

Do I need to study it to build a retrieval system? No. You can build one without it. You will have trouble improving one without it, because improvement requires measuring relevance, and measuring relevance is the field’s central method.

Where does “information architecture” fit? Information architecture applies the field’s organizing principles to websites and software interfaces. Richard Saul Wurman coined the term in 1976. Louis Rosenfeld and Peter Morville, both trained in library and information science, made it a web discipline with their 1998 O’Reilly book.

Where to go next

Start with the reference map of the AI stack, which places each of these concerns on one diagram. Then pick the topic closest to the problem you have: AI Architecture if retrieval returns the wrong passages, Ontologies if the agent confuses one entity with another, Knowledge Graphs if nobody can say where an answer came from.

• AI Architecture
• Ontologies
• Knowledge Graphs
• AI Infrastructure
• Data Modeling
• Intelligent Agents