Semantic Similarity
Updated July 21, 2026
Measuring how close two pieces of text are in meaning, even if they use different words. Used to determine whether a page change is substantive or cosmetic.
Also known as: Semantic Textual Similarity (STS), Semantic distance, Embedding similarity, Vector similarity, Meaning similarity
Semantic similarity measures how close two pieces of text are in meaning, independent of the exact words they use. Two sentences can share almost no vocabulary and still be near-identical in meaning ("cut prices by 20 percent" versus "slashed rates a fifth"), while two sentences that reuse the same words can mean different things. A semantic-similarity score captures that meaning-level closeness on a continuous scale, which is what makes it useful for deciding whether a competitor's page actually changed or was merely reworded.
The concept is old and well documented in computational linguistics and information retrieval. Its vector-space lineage traces to Gerard Salton's 1971 model, in which documents and queries are represented as vectors and compared by the cosine of the angle between them. That idea fed into distributional methods such as Latent Semantic Analysis in the early 1990s and, later, neural embeddings like word2vec and BERT. A parallel lineage computes similarity as distance between concepts in a lexical taxonomy such as WordNet, with foundational work by Resnik (1995), Jiang and Conrath (1997), and Lin (1998).
Today semantic similarity underpins search ranking, deduplication, plagiarism detection, biomedical ontology matching, and modern retrieval-augmented generation. In competitive-intelligence tooling it plays a narrower but pointed role: comparing the old and new text of a monitored page to separate substantive edits, like a price or feature claim, from cosmetic churn like whitespace, boilerplate, or minor rephrasing.
How semantic similarity is computed
The dominant modern approach converts text into embeddings, numeric vectors positioned so that passages with similar meaning land near each other in a high-dimensional space. Word-level embeddings such as word2vec or GloVe capture individual terms; sentence-level models such as Sentence-BERT or OpenAI-style embedding models capture whole sentences or content blocks. Once two texts are embedded, a distance metric scores their closeness.
Cosine similarity is the usual choice. It measures the angle between the two vectors rather than their length, so it reads directional alignment of meaning and ignores raw magnitude. Scores run from roughly zero, no meaningful overlap, to one, effectively equivalent. Many pipelines convert this to a distance by taking one minus the cosine value.
An older family skips embeddings entirely and walks a taxonomy like WordNet, scoring similarity from the path length or shared information content between two concepts. This works well for single word pairs but does not scale to comparing arbitrary sentences the way embeddings do, which is why change-monitoring systems lean on the embedding approach.
Semantic similarity vs. lexical and string similarity
Lexical or string similarity, including Jaccard overlap, n-gram matching, and Levenshtein edit distance, compares surface characters and words. It is fast and deterministic but blind to meaning. Reword a sentence completely and edit distance reports a large change even though nothing substantive shifted; keep most of the words but swap one number and it may report almost no change at all.
Semantic similarity inverts those failure modes. A paraphrased headline scores as highly similar because the meaning held, while a changed price or a flipped feature claim scores as dissimilar because the meaning moved, even if only a few characters differ. For competitor tracking this is the whole point: the interesting events are meaning changes, not character changes.
The two are complementary rather than rival. Change-detection systems often run cheap lexical or DOM-level diffing first to locate what changed, then apply semantic similarity to the changed blocks to judge whether the change matters.
How competitive-intelligence tools use it
A monitoring system crawls a competitor's page repeatedly and must answer one question on each pass: did anything meaningful change, or is this just markup churn. Alerting on every whitespace tweak, timestamp, rotating ad tag, or cosmetic rephrase floods users and trains them to ignore notifications.
Semantic similarity supplies the filter. The system embeds the previous and current versions of a page, or of individual content blocks, and computes their similarity. A high score means the meaning held and the change can be suppressed or downgraded; a low score means the meaning shifted and the change is worth surfacing, for example a new price, a repositioned value proposition, or an added feature claim. A threshold separates the two, and it is usually layered on top of standard HTML or DOM diffing so that trivial structural edits never reach the scoring stage.
In meertrack this is one input into deciding whether a detected change earns an alert. It pairs naturally with change-significance scoring and noise filtering, which weigh the semantic signal alongside where on the page the change occurred and what kind of content it touched.
Similarity vs. relatedness, and other limits
Strictly, semantic similarity is narrower than semantic relatedness. Similarity captures taxonomic, is-a closeness: car and bus are similar because both are vehicles. Relatedness covers any meaningful association, including function or cause and effect: coffee and mug are related but not similar. Many embedding models blur this line and score both, which is fine for change detection but matters when interpreting what a score claims.
The method has practical limits. Scores are only as good as the embedding model, and a model trained on general text may miss domain jargon, so a meaningful shift in niche pricing language can score as similar. Thresholds are not universal; the cutoff that works for prose can misjudge tables, code, or navigation. And embeddings can be fooled by negation or small numeric edits, where one flipped word or figure changes meaning more than the vector distance suggests. Treat the score as a strong prior, not a verdict.
Stop looking terms up. Start tracking them.
meertrack watches your competitors' websites, pricing, and hiring, then alerts you when something meaningful changes.
Frequently Asked Questions
What is semantic similarity in NLP?
It is a measure of how close two pieces of text are in meaning, regardless of the words they use. Systems typically map each text to a numeric vector called an embedding, then measure the closeness of those vectors. The result is a continuous score, from no shared meaning up to effectively equivalent, that lets software compare passages the way string matching cannot.
How is semantic similarity calculated?
The common method embeds each text as a vector using a model such as word2vec, GloVe, or Sentence-BERT, then scores the two vectors with cosine similarity, the cosine of the angle between them. Cosine is preferred because it measures directional alignment of meaning and ignores vector length. An older alternative computes similarity from the distance between concepts in a taxonomy like WordNet.
What is the difference between semantic similarity and lexical similarity?
Lexical or string similarity, such as Jaccard overlap or Levenshtein edit distance, compares surface characters and words and ignores meaning. Semantic similarity compares meaning. A paraphrase scores low on lexical similarity but high on semantic similarity, while swapping a single price digit barely moves lexical similarity but can sharply drop semantic similarity. Change-detection tools often use both in sequence.
How is semantic similarity used in website change detection?
A monitoring tool embeds the old and new text of a tracked page and computes their similarity. A high score means the meaning is unchanged, so cosmetic edits like rewording or whitespace get suppressed; a low score means the meaning moved, so the change is surfaced as a potential price, feature, or positioning update. A threshold separates substantive from cosmetic, usually on top of DOM diffing.
Is cosine similarity the same as semantic similarity?
No. Semantic similarity is the concept, how close two texts are in meaning. Cosine similarity is one mathematical metric, the angle between two vectors, commonly used to compute semantic similarity once text has been embedded. The two are often conflated in casual use, but cosine similarity is just one implementation choice; other distance measures can compute the same underlying concept.
Related terms
Assigning a relevance/importance score to each detected change using ML, so high-impact changes surface first.
Website Change DetectionAutomated monitoring of web pages to identify when content, structure, or visual appearance changes. The core technology underlying CI monitoring tools.
Noise FilteringAutomated suppression of irrelevant changes using rules, heuristics, or ML models. The key differentiator between "dumb" change detection and intelligent CI.
DOM DiffingComparing the Document Object Model (HTML structure) of a web page across two points in time to identify additions, removals, and modifications.
Natural Language Processing (NLP)The AI field focused on enabling computers to understand, interpret, and generate human language. Foundational for turning raw web text into actionable intelligence.
Large Language Model (LLM)A neural network trained on massive text corpora (GPT-4, Claude) capable of summarizing changes and generating human-readable alerts from raw data.
Retrieval-Augmented Generation (RAG)Combining a search/retrieval step with an LLM to generate responses grounded in specific documents. Used in CI tools for natural-language queries about competitor history.
Sentiment AnalysisClassifying text as positive, negative, or neutral. Applied in CI to gauge market reaction to competitor announcements and product reviews.