Does the model always return a correct analogy?

No; success depends on representations, relations, and evaluation rules.

Is it limited to two dimensions?

No; the arithmetic works in higher-dimensional vector spaces.

Is a nearest neighbor automatically analogous?

No. Similarity and relational correctness are distinct properties.

AI Glossary / Statistics and probability / Intermediate

Parallelogram model

Definition

The parallelogram model is a vector-based approach to analogy: it represents a relation as the difference between two vectors and applies that difference to a third vector. For an analogy A is to B as C is to what, it forms the target B - A + C and searches for a suitable nearby candidate.

What it means

The idea is that similar relationships may correspond to similar directions and distances in a representation space. If moving from A to B expresses a relation, moving the same way from C may point toward an analogous item.
In language processing, the items are often words represented by embeddings. The model is a geometric hypothesis about relations in those vectors. It does not mean that every semantic relation is linear or that successful vector arithmetic establishes human-like reasoning.

How it works

Choose representations that place the items in the same vector space. Mixing vectors from independently trained spaces generally makes subtraction meaningless unless a justified alignment has been performed.
Compute the relation offset B - A. Add it to C to obtain a predicted target. In exact geometry, A, B, C, and that target form the relevant parallelogram: the displacement from A to B equals the displacement from C to the target.
Search the candidate vocabulary or item set using a stated distance or similarity rule. Euclidean distance compares locations, while cosine similarity compares directions. Normalization choices affect results, so record them rather than treating all analogy implementations as identical.
Define candidate exclusions explicitly. The original query words or their variants can be close to the target and dominate retrieval. Excluding them changes the task and should be part of the evaluation protocol.
Evaluate multiple relations, not just memorable examples. Performance can depend on word frequency, ambiguity, neighborhood structure, and the relationship being tested. A system may exploit general similarity rather than accurately represent the intended relation.
Static word embeddings assign one vector per vocabulary item and can mix senses. Contextual representations require additional decisions about the sentence context, token position, and pooling. An analogy method that works on one representation need not transfer unchanged to another.
Treat retrieved answers as candidates. Validate the relation independently, especially when using domain vocabulary or making consequential decisions. The geometry can reflect corpus associations, including misleading patterns, rather than verified facts.

Three practical examples

Embedding evaluation: researchers test whether a representation supports a set of prespecified relational analogies.
Terminology exploration: a team explores candidate relations among domain terms, then asks subject specialists to verify them.
Education: an instructor uses a small vector example to show how relational offsets differ from simple nearest-neighbor similarity.

Think of it this way

Think of following an arrow on a map. You measure the arrow from one location to another and place the same arrow at a third location. Its endpoint suggests the fourth location. The analogy is useful only when that same movement actually represents the relevant relationship.

Worked example

Use an original toy space with A = (1,2), B = (3,3), and C = (2,5). These coordinates are invented to demonstrate the calculation; they are not learned word embeddings.
The relation offset is B - A = (2,1). Adding it to C gives target (4,6). Suppose the eligible candidate vectors are D = (4.1,5.9), E = (3,6), and F = (5,5).
Using Euclidean distance, the squared distances from the target are 0.1 squared + (-0.1) squared = 0.02 for D, 1 for E, and 2 for F. Their distances are approximately 0.1414, 1, and 1.4142, respectively. D is therefore the nearest eligible candidate.
The target itself forms an exact parallelogram, while D only approximately completes it. The error vector from the target to D is (0.1,-0.1).
This calculation establishes a ranking under a specified metric and candidate set. It does not establish that D has the intended semantic relationship. If an excluded query item or an additional closer candidate were included, the returned answer could change. Candidate selection is part of the procedure, not an incidental detail.

Illustrative case study

Illustrative fictional case: Moss Knowledge explores relationships in an embedding model trained on technical manuals. An analyst presents several successful analogies as evidence of broad relational understanding.
The review team adds less frequent terms, ambiguous terms, and relations from a held-out evaluation list. It checks whether retrieved candidates express the intended relation rather than merely sharing a topic. The team describes the resulting tool as an exploratory aid and retains expert verification. This fictional case illustrates evaluation discipline rather than a measured deployment outcome.

Common misconceptions and mistakes

Do not assume that geometric proximity proves a fact or relationship. Do not generalize from a handful of attractive analogy examples. A vector offset is not automatically a unique semantic direction, and different relations can overlap in the space.
Avoid silently changing normalization, similarity metrics, or query-word exclusions between experiments. Do not infer that an AI system uses human reasoning mechanisms merely because its answer matches a human response. Also avoid interpreting a two-dimensional projection as faithful evidence about every high-dimensional distance.

Business use

For business teams, the model is useful mainly as an explanation of embedding geometry and a possible exploratory technique. It can help diagnose representations or suggest terminology candidates. Operational use requires evaluation on relevant relations, transparent retrieval rules, and verification of the suggested answers against trusted domain knowledge.

Check your understanding

What is the target in the worked example?

Answer: B - A + C = (4,6).

Why does D win?

Answer: Its Euclidean distance to that target is smallest among the stated candidates.

What remains unproven?

Answer: Whether the candidate satisfies a meaningful real-world relation.

Related terms

Learn next

Next compare similarity retrieval with relational analogy evaluation. Then examine normalization, candidate exclusions, and failure cases across different relation types and representation models.

Sources

Daniel Jurafsky and James H. Martin, Speech and Language Processing, Chapter 5, Embeddings, August 19, 2026 draft: https://web.stanford.edu/~jurafsky/slp3/5.pdf — author-hosted discussion of the parallelogram method and its limitations. The draft's analogy discussion was checked against primary search excerpts. Coordinates, distance calculations, and the fictional case are original.

Previous: Parallelization contract Next: Parameter