Open the vocabulary section of almost any language app and you will see the same thing: a list. Words in one column, translations in another, maybe a strength bar. It is a spreadsheet view of something that, inside your head, looks nothing like a spreadsheet. This article covers what the research actually says about how words are stored, the two practical consequences that matter for learners, and how we have started building LingoBlend around that shape.
What the mental lexicon actually looks like#
The term psycholinguists use for your word store is the mental lexicon, and the classic model of it is a semantic network: words as nodes, relationships as connections, and activation spreading along those connections (Collins & Loftus, 1975). When you hear doctor, the words nurse, hospital, and medicine become measurably faster to recognize for the next fraction of a second — an effect called semantic priming, and one of the most replicated findings in the field.
The connections are not only about meaning. Word-association studies — say a word, record the first word that comes back — show stable, structured links of several kinds:
| Connection type | Example (Spanish) | What it reflects |
|---|---|---|
| Word family | comer → comida, comedor | Shared root, morphology |
| Synonym / antonym | feliz → contento, triste | Meaning overlap and contrast |
| Topic / collocation | comer → cocina, almuerzo | Words that live in the same situations |
| Form similarity | pero ↔ perro | Sound- and spelling-alikes |
Second-language research adds the finding that matters most for learners: network density tracks proficiency. Paul Meara's work on L2 lexical networks argued that vocabulary size alone is a weak measure — beginners' word stores behave like loose collections, while advanced learners' behave like native-style webs, with associations shifting from sound-based ("what does it rhyme with?") toward meaning-based as proficiency grows. Two learners can know the same 2,000 words, and the one whose words are connected reads faster, recalls faster, and guesses new words better.
Consequence one: new words need something to attach to#
A new word that connects to things you already know has somewhere to live. This is the vocabulary version of the comprehensible input principle — learning happens at the edge of what you know, not in a vacuum. If you already know comer (to eat), then comida (food) is barely a new word at all: it inherits the root, the topic, and half its contexts from a node you already own. Learning it costs a fraction of what an unrelated word costs.
That edge — the set of words one connection away from your known vocabulary — is your personal frontier, and it is where the cheapest learning lives. Frequency lists ignore it entirely: they hand every learner the same next word, regardless of what that learner already knows. A network model can rank the frontier for you specifically. (This is not a new idea in learning software generally — math platforms have used prerequisite graphs to pick "what's next" for years — but it is rare in language apps.)
Consequence two: similar words fight each other#
The second prediction is less intuitive and better supported: studying similar words together makes them harder to learn. A line of research going back decades (Tinkham, 1993; Waring, 1997) found that learning semantically clustered sets — all the colors at once, all the fruit at once — is measurably slower than learning the same words in mixed sets, because the similar items interfere with each other. Every learner who has permanently cross-wired Tuesday/Thursday or Spanish pero/perro knows the feeling. Paul Nation's vocabulary-teaching guidelines say it plainly: present related and confusable items at different times.
The practical upshot: an app should know which of your words are confusable — and use that knowledge to keep them apart when they are new, and to contrast them deliberately once you keep mixing them up. You can only do either if the confusion pairs are mapped.
What this means for how an app should model you#
Put the two consequences together and the requirements fall out naturally:
| Flat-list model | Network model |
|---|---|
| Knows which words you saved | Knows which words you saved and how they relate |
| Next word = next row in a frequency list | Next word = your frontier, ranked |
| Blind to confusables | Can separate pero/perro in time, then contrast them |
| despierta and despertar are two unrelated rows | Both are the same word family, and knowing one credits the other |
| Vocabulary size as the only metric | Size and connectedness |
That last row deserves a note, because it is where list-model apps quietly leak your progress. Learners meet words in inflected forms — despierta, not the dictionary form despertar — and a list model stores whatever form you saved as its own isolated row. A network model resolves forms to their family, so the knowledge accumulates in one place.
How LingoBlend builds this#
We have now built a vocabulary graph for every one of our 17 languages: roughly 3,000 of the most useful words per language — selected from real usage frequency, not textbook order — connected by word families, synonyms, antonyms, confusables, topics, and semantic similarity, plus an inflection index that maps tens of thousands of surface forms to their dictionary forms (despierta → despertar, and irregulars a spell-pattern rule could never catch, like Italian va → andare).
Your side of the graph comes from what LingoBlend already tracks: every word you save, and how your reviews go in the spaced-repetition games. That review history classifies each of your words — new, learning, young, mastered, struggling — and lays your knowledge over the language's map. The computation happens on your device, costs nothing, and works offline.
Today this runs under the hood. It is the foundation for the personalization we are rolling out next: blends that stop spending translation slots on words you have already mastered, recommendations drawn from your personal frontier instead of a generic list, and review sessions that respect the interference research above. If you want the deeper technical tour of what is in the graph, we wrote one: inside LingoBlend's vocabulary graph.
Related reading: how to build vocabulary by reading for the coverage and repeated-encounter research, the science of bilingual reading for why mixed-language text holds attention, and measure your vocabulary size for what size does and does not tell you. Or see the full research grounding on our science page.