Is ESA a type of word embedding?

It creates semantic vector representations, but its defining feature is explicit knowledge-based concept dimensions; it should not be treated as synonymous with learned dense embeddings.

Must it use Wikipedia?

Wikipedia is central to the original work, but the authors also explored another concept collection.

Does it remove word ambiguity?

Contextual contributions can help favor relevant concepts, but ambiguous or poorly covered input can still produce misleading associations.

AI Glossary / Natural language and speech / Intermediate

Explicit semantic analysis

Definition

Explicit semantic analysis, usually abbreviated ESA, represents text as a weighted vector of named concepts drawn from a knowledge collection. In the original approach, encyclopedia articles supply those concepts. Texts can then be compared through their concept vectors, including when they use different words to discuss related subjects.

What it means

A concept dimension is associated with an identifiable article or category rather than an unnamed latent coordinate. For example, a text might receive weights for concepts associated with gardening, irrigation, and agriculture.
The word explicit refers to those identifiable dimensions. It does not mean that every association is correct, that the vector is a complete explanation, or that the method understands the text as a person does.
ESA is particularly associated with the work of Evgeniy Gabrilovich and Shaul Markovitch. Their research explored Wikipedia and another human-organized knowledge collection. The general representation should therefore not be confused with a requirement to use one fixed encyclopedia snapshot.

How it works

Prepare a concept collection in which each concept has descriptive text. Process the words and calculate weights, such as term frequency - inverse document frequency, to represent each concept's description.
Create an inverted index that connects a word to concepts whose descriptions contain it, retaining their association weights. A new input text supplies its own weighted words.
Combine the word-to-concept associations according to the input word weights. In a simple weighted sum, each concept receives contributions from all the relevant input words.
Use the resulting vector to compare texts, often through cosine similarity, or as features for another task. Preprocessing, weight normalization, pruning, and the concept collection affect the output, so an implementation should document these choices.

Three practical examples

A research search system compares a short question with documents through shared concept associations, supplementing exact word overlap.
A text-classification experiment adds concept features to ordinary word features and tests whether they help distinguish its categories.
An analyst inspects highly weighted concepts for two descriptions to understand which reference topics connect them, while checking whether those topics actually fit the context.

Think of it this way

Think of placing a document beside labeled reference shelves and recording how strongly it connects to each shelf. Two documents can connect to similar shelves without sharing many words. The analogy stops where a librarian's judgment begins: ESA's associations come from its representations and weighting procedure, not a human assessment of each new document.

Worked example

Consider an invented three-concept collection with dimensions Gardening, Cooking, and Astronomy. Suppose the association vector for the word “plant” is (0.8, 0.1, 0), and for “water” it is (0.6, 0.3, 0.1).
For a short input with word weights 2 for plant and 1 for water, a weighted-sum representation is 2 × (0.8, 0.1, 0) + (0.6, 0.3, 0.1) = (2.2, 0.5, 0.1).
Gardening receives the largest weight, but Cooking and Astronomy remain nonzero. These values are associations, not probabilities, and their sum need not equal one.
If another input has the same word proportions but all weights doubled, its vector becomes (4.4, 1.0, 0.2). Its cosine similarity with the first is one because the vectors point in the same direction. That fact reflects this representation, not proof that the texts have identical meaning.
All numbers and concept labels here are constructed for arithmetic practice, not extracted from Wikipedia or a deployed ESA system.

Illustrative case study

In a fictional project, Birch Knowledge compares short customer questions with a collection of educational articles. Some questions use everyday expressions while the articles use technical vocabulary.
The team builds an ESA experiment and examines the top concepts for successful and unsuccessful matches. One ambiguous word activates an unrelated topic, so reviewers add that case to a test set and inspect how surrounding words influence the vector.
The imagined project demonstrates a practical inspection process. It claims neither a measured search improvement nor that named concept dimensions automatically resolve ambiguity.

Common misconceptions and mistakes

Explicit concepts are not the same as manually verified labels assigned to each new text.
ESA is not identical to latent semantic analysis, whose dimensions arise from a different transformation of a term-document representation.
A high similarity score indicates closeness under the chosen representation, not necessarily synonymy, factual agreement, or an answer to the same question.
An encyclopedia's coverage, wording, and historical snapshot can influence the associations. Human authorship does not make a collection free of omissions or bias.
More concept dimensions do not by themselves guarantee better results; irrelevant associations and computational cost still matter.

Business use

For a business knowledge-search experiment, ESA offers a way to connect text with identifiable reference topics. Its value depends on whether the reference collection covers the organization's vocabulary and whether concept similarity improves the actual retrieval task.
Compare it with a clear baseline, inspect ambiguous and specialized queries, and record the collection version. Present concept weights as diagnostic evidence rather than a guarantee that a retrieved answer is correct.

Check your understanding

In the worked example, does a Gardening weight of 2.2 mean a 220 percent probability that the input concerns gardening?

Answer: No. It is a weighted association obtained by addition. It is not a calibrated probability, and the representation's scale depends on the input and weighting rules.

Related terms

Learn next

calculate cosine similarity for two small concept vectors, then compare which shared words and named concepts contribute to their relationship. Study latent semantic analysis afterward to understand how explicit reference dimensions differ from latent dimensions.

Sources

Gabrilovich and Markovitch, “Computing Semantic Relatedness using Wikipedia-based Explicit Semantic Analysis,” original IJCAI 2007 paper: https://www.ijcai.org/Proceedings/07/Papers/259.pdf. The paper explains concept vectors, weighted interpretation, semantic comparison, and experiments; the numerical example above is original teaching material. Independent source review: the original paper’s section 2 weighted concept-vector construction and section 3 alternative Open Directory Project implementation were read.

Previous: Explicit parallelism Next: Exploding gradient problem