When Close Enough Isn't: The Gap Between Vector Similarity and Genuine Language Comprehension
There is a quiet assumption embedded in nearly every modern NLP pipeline: that proximity in vector space corresponds to proximity in meaning. It is an assumption so thoroughly baked into how we evaluate embeddings, design retrieval systems, and benchmark models that questioning it can feel almost heretical. Yet the assumption is wrong often enough—and wrong in consequential enough ways—that every engineer working with language models should understand precisely where it breaks down.
This is not an argument against embeddings. Dense vector representations remain among the most powerful tools available for applied NLP. It is, rather, an argument for treating cosine similarity as a measurement of statistical association rather than semantic equivalence, and for designing systems that account for the difference.
The Geometry of Meaning: What Embeddings Actually Capture
When a transformer-based encoder maps a sentence into a high-dimensional vector, it is not performing some form of linguistic understanding in the way humans do. It is learning, through exposure to vast corpora, which words and phrases tend to appear in similar distributional contexts. The distributional hypothesis—the idea that words appearing in similar contexts tend to have similar meanings—is the theoretical backbone of the entire field.
The problem is that distributional similarity and semantic similarity are related but not identical. Consider the pair "bank" (financial institution) and "bank" (riverbank). In many embedding spaces, these will land in different regions because their contexts differ. That is the system working as intended. But now consider "cheap" and "affordable." Their cosine similarity will be high—legitimately so. However, consider "cheap" and "inexpensive hotel" versus "cheap" and "low-quality product." Both may score high similarity against a query about "budget accommodations," yet only one reflects what the user actually wants.
This is the first failure mode: polysemy collapse and connotation blindness. Embeddings trained on general corpora flatten nuances of register, connotation, and pragmatic context into a single vector. The result is a space where synonyms, near-synonyms, and associated-but-distinct concepts occupy overlapping neighborhoods.
Shortcuts in the Training Data
The deeper issue lies in how models generalize from training data. Neural networks are extraordinarily good at finding shortcuts—statistical patterns that predict the correct answer on training examples without encoding the underlying principle. In NLP, this manifests as spurious correlations that inflate similarity scores without grounding them in genuine semantic relationships.
A concrete example: in many general-purpose embedding models, the phrases "not safe" and "safe" will have higher cosine similarity than intuition would suggest. The reason is that both phrases appear in similar contexts—safety discussions, product descriptions, regulatory documents—even though their meanings are antonyms. The model has learned contextual association, not logical opposition.
This phenomenon extends to cross-domain retrieval. A semantic search system trained on web text and deployed over a medical knowledge base may surface results that are distributionally similar to a query without being medically relevant. The phrase "cardiac event" and "corporate event" share enough surface-level context in training data that their embeddings may be closer than expected, creating retrieval failures that are difficult to diagnose precisely because the similarity score looks reasonable.
Cross-Lingual Transfer and the Illusion of Alignment
Multilingual embedding models add another layer of complexity. Models like mBERT and multilingual variants of sentence transformers claim to align semantic content across languages by mapping equivalent meanings to nearby points in a shared vector space. In practice, this alignment is uneven and often reflects typological similarity between languages rather than semantic equivalence.
High-resource language pairs—English and French, for instance—tend to align well for common vocabulary and standard sentence structures. But the alignment degrades sharply for idiomatic expressions, culture-specific concepts, and low-frequency vocabulary. A query in Spanish about a regional legal concept may retrieve English documents about a superficially similar but legally distinct concept, because the embedding model has no mechanism for distinguishing between semantic equivalence and translational approximation.
For organizations deploying multilingual retrieval or cross-lingual recommendation systems at scale, this is not a theoretical concern. It is a source of measurable accuracy degradation that cosine similarity scores will not reveal.
Diagnosing Shortcut Learning in Your Embedding Space
The good news is that shortcut learning in embedding spaces is diagnosable, even without access to model internals. The following approaches offer practical entry points for engineering teams.
Contrastive probing sets. Construct evaluation sets where the correct answer requires distinguishing between semantically similar but meaningfully distinct items—antonyms, near-synonyms with different connotations, domain-specific terms with general-corpus counterparts. If your embedding model cannot reliably rank the correct item above plausible distractors, you have evidence of shortcut reliance.
Nearest-neighbor auditing. For a representative sample of your query or document corpus, retrieve the top-k nearest neighbors and manually assess whether the retrieved items are genuinely semantically related or merely contextually associated. Pay particular attention to cases where high similarity scores correspond to topically adjacent but logically distinct content.
Sensitivity analysis on negation and qualification. Test whether your embedding model treats negated or heavily qualified statements differently from their affirmative counterparts. Systematic insensitivity to negation is a reliable indicator that the model is encoding topic rather than meaning.
Cross-domain transfer evaluation. If your production system operates in a specialized domain—legal, medical, financial—evaluate embedding performance on domain-specific benchmarks rather than relying on general-purpose leaderboard scores. Models that rank highly on MTEB or similar benchmarks may underperform significantly on domain-specific semantic tasks.
Engineering Around the Gap
Understanding the limitation is only the first step. Several practical strategies can mitigate the gap between vector similarity and semantic accuracy in production systems.
Fine-tuning on domain-specific contrastive data remains the most reliable approach. Training a bi-encoder or cross-encoder on pairs drawn from your actual deployment domain—with hard negatives that reflect the specific failure modes you have identified—substantially improves the alignment between similarity scores and genuine relevance.
Hybrid retrieval architectures that combine dense vector search with sparse lexical signals (BM25 or similar) provide a natural check on embedding shortcuts. Lexical signals are less susceptible to distributional conflation, and cases where the two retrieval modes disagree are often precisely the cases where the embedding model is relying on spurious correlations.
For high-stakes applications, reranking with a cross-encoder adds a second inference step that considers query and candidate jointly, rather than as independent vectors. Cross-encoders are slower and more expensive, but they are substantially better at capturing fine-grained semantic distinctions that bi-encoder similarity scores miss.
The Measurement Problem
Perhaps the most important takeaway for engineering teams is this: cosine similarity is a measurement of geometric proximity, not a measurement of meaning. The two correlate—sometimes strongly—but the correlation is neither universal nor reliable enough to serve as the sole quality signal for a production semantic system.
Building robust NLP infrastructure requires acknowledging that the vector space is a model of language, not language itself. The shortcut between distributional similarity and semantic understanding is convenient, but it is also the source of some of the most persistent and difficult-to-diagnose failures in applied NLP. Treating it with appropriate skepticism is not pessimism about the technology—it is the foundation of engineering systems that actually work.
