Why do embeddings measure semantic association instead of task relevance?
This explores why a tool meant to find the *right* information (task relevance) instead finds merely *related* information (semantic association) — and what that gap costs in real systems.
This explores why embeddings find things that are *related* to your query rather than things that actually *answer* it — a gap that's invisible in demos but punishing in production. The root cause is mechanical: embeddings are built from co-occurrence statistics, so they encode which words and concepts tend to appear together, not which ones play the role your task needs. The clearest statement of the problem is that vector embeddings measure semantic association, not task relevance Do vector embeddings actually measure task relevance? — two concepts that are semantically close but play different roles (say, a symptom vs. a treatment) land near each other, so an underspecified query surfaces a crowd of wrong-but-associated candidates.
Why is association baked in so deeply? Because that's literally what the geometry is built from. Analysis of embedding spaces shows their leading directions split a taxonomy coarse-to-fine, mirroring the WordNet hypernym tree level by level — a structure that falls directly out of co-occurrence statistics Do embedding eigenvectors organize taxonomy from coarse to fine?. And static embeddings really do carry genuine meaning — clustering reveals sensitivity to valence, concreteness, and other psycholinguistic measures before attention even operates Do transformer static embeddings actually encode semantic meaning?. So the embedding isn't broken; it's faithfully measuring the thing it was trained to measure. Task relevance is simply a *different* thing that nobody encoded.
What's worth noticing is that this 'association beats intent' pattern is not unique to retrieval — it's the same failure showing up across the corpus under different names. Language models prefer high-frequency surface phrasings over semantically equivalent rare ones, tracking statistical mass from pretraining rather than meaning Do language models really understand meaning or just surface frequency?. They reason through semantic association rather than symbolic logic, collapsing when meaning is decoupled from the rules Do large language models reason symbolically or semantically?. And they ignore in-context information when prior training associations are strong enough to override it Why do language models ignore information in their context?. Embeddings-measure-association is the retrieval-layer version of a system-wide bias: these models default to 'what usually goes together' over 'what this specific situation requires.'
The interesting part is what people do about it — and the fixes converge on the same move: *break the tight coupling to raw text similarity.* In recommendation, VQ-Rec quantizes item text into discrete codes that index learned embeddings, deliberately decoupling representation from text so similarity bias stops leaking into results Can discretizing text embeddings improve recommendation transfer?. In vision, SignRAG sidesteps direct embedding similarity entirely — it describes an image in natural language, then retrieves against a text index, because the description bridges the reference gap better than embedding distance does Can describing images in text improve zero-shot recognition?. Both treat the embedding's native semantic geometry as the problem to route around, not the answer.
The takeaway a reader might not expect: 'relevance' isn't a property of the data, it's a property of *your task* — and embeddings don't know your task. They give you a general-purpose map of what's near what. Turning that into task relevance means adding the missing ingredient, whether that's an intermediate code, a natural-language description, or a downstream model that learns the role you actually care about.
Sources 8 notes
Embeddings encode co-occurrence patterns, making semantically close but role-distinct concepts highly similar. This works in simple demos but fails in production where underspecified queries have many wrong-but-associated candidates.
Leading eigenvectors of embedding Gram matrices separate broad taxonomic branches first, then progressively finer sub-branches—a coarse-to-fine spectral order that tracks the WordNet hypernym tree level by level, confirming predictions from co-occurrence statistics.
Clustering analysis of RoBERTa embeddings reveals sensitivity to five psycholinguistic measures including valence, concreteness, iconicity, and taboo. This demonstrates that static embeddings function as genuine lexical entries containing semantic content before self-attention operates.
LLMs show consistent preference for higher-frequency surface forms over semantically equivalent rare paraphrases across math, machine translation, commonsense reasoning, and tool calling. This suggests models track statistical mass from pretraining rather than meaning-recognition as their primary mechanism.
When semantic content is decoupled from reasoning tasks, LLM performance collapses even with correct rules in context. Models rely on parametric commonsense and token associations rather than formal logical manipulation, constraining reasoning to training distribution semantics.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
VQ-Rec uses product quantization to map item text to discrete codes that index learned embeddings, breaking the tight coupling between text and recommendations. This decoupling prevents text-similarity bias and allows lookup tables to adapt to new domains without retraining the text encoder.
SignRAG demonstrates that describing an unknown image via vision-language model, then retrieving known designs from a text-indexed database, eliminates the need for recognition model training. Natural-language description bridges the visual-reference gap better than direct embedding similarity.