Can related terms have different parts of speech?

Yes. Connections can cross grammatical categories.

Does a cosine of one prove identical meaning?

No. It indicates identical direction for the vectors being compared.

Is relatedness always useful?

Its usefulness depends on what the application is trying to accomplish.

AI Glossary / Natural language and speech / Advanced

Relatedness

Definition

Relatedness is the degree to which words, concepts, or other items are meaningfully connected. In lexical semantics, it is broader than similarity: two words can be related through a shared activity, association, or functional relationship without meaning nearly the same thing. For example, a violin and a bow are related through musical use, but they are not interchangeable objects.

What it means

A system asked for related terms may return instruments, materials, actions, places, and opposites associated with the input. A system asked for synonyms has a narrower job. Confusing these requests can produce suggestions that appear relevant while changing the intended meaning.
Consider gardener and shovel. They belong to a familiar activity, but one names a person and the other a tool. Their connection can be useful when retrieving information about gardening. It does not justify replacing gardener with shovel in a sentence.
Relatedness also depends on the sense and context being considered. Mouse may connect to cheese in one setting and to keyboard in another. A score for the isolated spelling can mix associations that a particular user intended to keep separate.

How it works

Human judgments can provide relatedness ratings under explicit instructions. Computational systems can estimate relatedness from patterns of occurrence, learned representations, or links in a knowledge resource. Each method captures evidence shaped by its data and design.
For vector representations, cosine similarity is often used to compare directions. Despite its name, a high cosine value does not by itself prove synonymy. If the representations encode shared topics or contexts, the score may reflect association rather than substitutability. Interpreting the number requires knowing what the representation was trained to preserve.
A knowledge-based approach may identify explicit connections such as used-for or part-of. A broad relatedness score can summarize that an association exists while losing the type and direction of the relationship. For an application that needs to know who uses what, a scalar score alone is insufficient.
Evaluation should therefore match the intended relation. A benchmark designed for similarity is not automatically a benchmark for every kind of association. SimLex-999 was created specifically to separate similarity judgments from broader relatedness or association, illustrating why the distinction matters when assessing semantic models.

Three practical examples

A search tool expands a query about bicycle maintenance with associated concepts such as tires and pumps, then checks whether the additional results serve the user's need.
A writing assistant distinguishes related vocabulary from actual paraphrase candidates so that an associated word is not presented as a synonym.
A topic-exploration interface groups connected concepts while labeling the relationship where evidence supports a more specific description.

Think of it this way

Think of relatedness as belonging on the same workshop wall. A hammer, a nail, and a repair instruction may all belong there together. That does not mean they perform the same role or can replace each other. The wall shows connection; a more detailed account explains the connection's nature.

Worked example

Consider an original toy representation with two coordinates: occurrence in gardening contexts and occurrence in music contexts. These invented vectors illustrate arithmetic, not a trained model.
Let gardener=[4,0], shovel=[3,0], violin=[0,5], and bow=[0,2]. For this exercise, bow is restricted to the musical sense. Cosine is the dot product divided by the product of vector lengths.
The gardener - shovel dot product is 4×3+0×0=12. Their lengths are four and three, so cosine is 12/(4×3)=1. The violin - bow dot product is ten, and their lengths are five and two, giving 10/(5×2)=1.
The gardener - violin dot product is zero, producing cosine zero. In this representation, the within-activity pairs have identical directions even though their magnitudes differ. That is enough for maximum cosine, but neither pair consists of synonyms.
Suppose a hypothetical suggestion rule labels every pair with cosine above 0.9 as interchangeable. It would label both gardener - shovel and violin - bow that way. Under the intended senses, both decisions are wrong. The error lies in interpreting the score.
Now add a hypothetical mixed-context word with vector [1,1]. Its cosine with gardener is 4/(4×√2)=1/√2, approximately 0.7071. The same score results with violin, reflecting balanced context coordinates rather than identical meaning relationships.
The example also exposes a limitation: two context dimensions cannot express all the distinctions a real language application needs. More dimensions may help represent information, but dimension count alone does not guarantee that a model distinguishes association, similarity, opposition, and identity correctly.

Illustrative case study

Fictional example: WordPath builds a vocabulary assistant from a model originally used for topic retrieval. Users asking for alternative wording receive suggestions that are associated with the topic but change the sentence's meaning.
The team separates an “explore related terms” feature from its replacement suggestions. It evaluates replacements in complete sentences and checks the intended word sense. Reviewers also distinguish useful associations from unsupported relationship labels. The scenario illustrates matching a semantic model to its use; it reports no measured improvement or recommendation of a particular model.

Common misconceptions and mistakes

Related words are not necessarily synonyms, and a strong association is not evidence of causation. Opposites may be closely connected because they occur in discussions of the same dimension. A numerical proximity score does not resolve which relationship is present.
Another mistake is treating relatedness as independent of context, language, and dataset. A corpus can emphasize specialized associations that are uncommon elsewhere. It is also misleading to compare scores from different models as though their numerical scales had identical meanings. Calibration and evaluation depend on the intended task.

Business use

Clarify whether a product needs broad discovery, near-synonym suggestions, entity matching, or a specific relation such as ownership. Use examples that separate those cases during evaluation. For search expansion, examine added irrelevant results as well as useful discoveries. For rewriting, check the complete sentence rather than relying solely on nearest neighbors in a vector space.

Check your understanding

Why do gardener and shovel have cosine one in the example?

Answer: Their vectors point along the same context axis.

Are they therefore interchangeable?

Answer: No. They name different kinds of things.

What does the mixed vector's equal score establish?

Answer: Equal directional proximity in this toy representation only.

Related terms

Learn next

Compare pairs that are synonymous, associated, opposite, or unrelated under specified senses. Define which relationships your application should reward before selecting its evaluation data.

Sources

Jurafsky and Martin, Speech and Language Processing, chapter 5, lexical semantics and vector representations: https://web.stanford.edu/~jurafsky/slp3/5.pdf Hill, Reichart, and Korhonen, SimLex-999: Evaluating Semantic Models With (Genuine) Similarity Estimation, 2015: https://aclanthology.org/J15-4004/

Previous: Rejection sampling Next: Relation