Latent semantic analysis¶

A distributional-semantic technique that applies truncated singular-value decomposition to a term–document matrix so terms and documents are represented in a shared lower-dimensional latent space.

Version: v2 · 2026-09-19 · History
Domain-specific #: 5267
Domain group: Applied Sciences & Engineering
Origin domain: Computer Science & Software Engineering
Subdomain: Natural Language Processing and Information Retrieval → Computer Science & Software Engineering

Core Idea¶

Weighting, rank, tokenization and corpus determine the geometry; LSA smooths synonymy and noise but can merge distinct senses and does not model word order or context dynamically. Counts or weighted term occurrences form a sparse matrix, SVD decomposes it into orthogonal factors, smaller singular directions are discarded and cosine or related similarity is computed among the reduced vectors. The abstraction is therefore identified by a declared carrier, a transformation or constraint over that carrier, and an invariant that tells an analyst whether the named structure is genuinely present.

How would you explain it like I'm…

Word Neighbor Squisher

Imagine a giant chart showing which words appear in which books. A computer squishes the chart down so that only its biggest patterns are kept. Words that show up in the same kinds of books end up close together, like 'car' and 'automobile'. But it doesn't notice what order the words come in.

Hidden Word Patterns

Latent semantic analysis starts by counting how often each word appears in each document, which makes a huge table that is mostly zeros. A math tool breaks the table into main patterns, and the weakest patterns are thrown away. Every word and document becomes a short list of numbers, and things with similar lists are treated as related. This helps match words that mean the same thing. It can also mix up a word that has two meanings, and it ignores word order.

Low-Rank Word–Document Similarity

Latent semantic analysis builds a term-document matrix of counts, often weighted (for example with tf-idf). It then applies singular value decomposition (SVD), which splits the matrix into independent factors ranked by importance. Keeping only the top factors gives a low-dimensional space in which words and documents are vectors, and similarity is usually measured by cosine. The reduction smooths over synonyms and noise. It can also blend distinct senses of a word, and it ignores word order and context. Results depend on choices like the corpus, tokenization, weighting and how many factors are kept, so those choices must be stated.

Latent semantic analysis builds a sparse term–document matrix of raw or weighted counts, such as TF-IDF or log-entropy weights. It computes a truncated singular value decomposition, keeping the k largest singular values and their orthogonal singular vectors. Terms and documents are then embedded in this rank-k space, and similarity is computed there, typically as cosine similarity. The truncation acts as a low-rank smoother, so co-occurrence structure can align synonyms and suppress noise. The resulting geometry is not intrinsic to the method. It depends on the corpus and document boundaries, the vocabulary and tokenization, matrix orientation, the weighting scheme, any centering, the SVD algorithm, and the retained rank. A full specification should also state how new documents are mapped into the space (folding-in), the similarity metric, the evaluation, the computational cost, and limits on interpretation. Known limitations include conflating distinct word senses into one vector and having no dynamic model of word order or context.

Scope of Application¶

Latent semantic analysis belongs to natural language processing and information retrieval and is useful where the analyst can specify the typed natural language processing and information retrieval carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, then evaluate the corpus and document boundary, vocabulary and tokenization, term-document orientation, count and weighting scheme, centering if any, SVD algorithm, retained rank, term and document embeddings, similarity metric, out-of-sample mapping, evaluation, computational cost and interpretability limits are explicit.

Clarity¶

The abstraction clarifies a crowded vocabulary by making the corpus and document boundary, vocabulary and tokenization, term-document orientation, count and weighting scheme, centering if any, SVD algorithm, retained rank, term and document embeddings, similarity metric, out-of-sample mapping, evaluation, computational cost and interpretability limits are explicit the center of the account. A claim should name the carrier, the governing operation or relation, the applicable assumptions, and the recognition test.

Manages Complexity¶

Without the abstraction, an analyst must reason directly over many local details: the carrier roles, admissibility assumptions, competing conventions, derived invariants, boundary cases, and proof or validation obligations specific to Latent semantic analysis. Latent semantic analysis compresses them into the roles in the structural signature. That compression permits comparison across instances without erasing the variables that determine validity. It also exposes which details may be varied safely and which are constitutive.

Abstract Reasoning¶

Identify the carrier. State what the elements, states, objects, or observations are: the typed natural language processing and information retrieval carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets. Reject examples whose alleged carrier belongs to a different problem. 2.

Knowledge Transfer¶

Knowledge transfers strongly among subfields of natural language processing and information retrieval because they reuse the typed natural language processing and information retrieval carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, Counts or weighted term occurrences form a sparse matrix, SVD decomposes it into orthogonal factors, smaller singular directions are discarded and cosine or related similarity is computed among the reduced vectors., and type the carrier, state every parameter and convention in the definition, test that the corpus and document boundary, vocabulary and tokenization, term-document orientation, count and weighting scheme, centering if any, SVD algorithm, retained rank, term and document embeddings, similarity metric, out-of-sample mapping, evaluation, computational cost and interpretability limits are explicit, compare the nearest accepted identity, and report counterexamples, uncertainty, and limiting cases.

Relationships to Other Abstractions¶

Current abstraction Latent semantic analysis Domain-specific

Parents (1) — outgoing typed relations

Latent semantic analysis is a kind of Dimensionality Reduction Prime

It remains its own entry because its identity is fixed by the corpus and document boundary, vocabulary and tokenization, term-document orientation, count and weighting scheme, centering if any, SVD algorithm, retained rank, term and document embeddings, similarity metric, out-of-sample mapping, evaluation, computational cost and interpretability limits.

Hierarchy paths (4) — routes to 3 parentless roots

Latent semantic analysis → is a kind of Dimensionality Reduction → is a kind of Approximation → is a decomposition of Representation → presupposes Abstraction

Show alternative paths (3)

Neighborhood in Abstraction Space¶

Latent semantic analysis sits in a crowded region of the domain-specific corpus (12^th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Lexical & Compositional Semantics (24 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-25