The Missing Infrastructure for Word Games

Vocabulary, definitions, associations, word families, and 1.46M usage examples in one licensable package.

Linguabase was built to power word games. It’s been in development since 2011 and currently powers “In Other Words”, a live iOS game where players navigate semantic space. Linguabase provides the data layer for anyone who wants to explore this space.

Linguabase is built on a foundation of human-generated data—professional lexicographic work, curated word lists, and structured linguistic resources accumulated over a decade. This foundation is then enhanced through focused LLM queries that validate, rank, and expand relationships. The result is an amalgam that neither human curation nor LLM inference could produce alone.

Three Things Every Word Game Needs

Words, links, meaning. License what you need:

Delivered as files you can embed (TSV, SQLite, JSON) or query via API. See delivery options →

Vocabulary

The vocabulary layer provides word lists with difficulty rankings, enabling you to adjust your vocabulary for the difficulty and obscurity level your game needs.

Why 400K Words?

In our experience, 400,000 words delivers an unabridged experience that is complete without being noisy. Larger word counts contaminate the player experience with obscure and borderline words. Smaller counts often miss legitimate words players expect to find. We can vary the delivered word count, but 400K is our recommendation.

For comparison:

Related
Words with Spaces — why “boiling water” and half a million other compound phrases aren’t in any dictionary.But these numbers obscure a deeper question: what counts as “one word”? Every dictionary draws the line differently, and it’s nearly impossible to count how many words are in English—even with clear inclusion policies on British vs. American spelling, proper nouns, accent variants, and punctuation. (For reference, Roget’s original 1852 thesaurus covered the working vocabulary of English with just 15,000 words across 1,000 semantic categories—a fraction of modern dictionary counts, yet considered comprehensive for its purpose.)

Consider these cases where reasonable people disagree:

Which of the above count as words? All of them? Just some? And which count as individual words?

Linguabase includes all spelling variants for common words and primary variants for less common ones—so your game recognizes “naïve” and “naive,” “café” and “cafe,” without inflating word counts artificially or rejecting spellings players reasonably expect to work.

Difficulty Rankings

Measuring usefulness for ranking is intuitive but difficult to quantify. Frequency of appearance in books or spoken word corpora would rank “the” and “of” highest while missing that “red panda” is easily understood by most English speakers. PageRank-style algorithms surface superconnectors, but that’s not the same as how understandable something is.

The Linguabase word ranking is an effective proxy for what words are known to a broad user base, including English language learners. The highest ranks align closely with vocabulary standards like the Oxford 3000 and CEFR levels (A1–B2)—but enhanced with common derivatives. Where Oxford 3000 lists “actor,” our ranking includes both “actor” and “actors.”

Here’s what you find at different ranks of the Linguabase ranking:

For “In Other Words,” we thresholded at ~400K. This includes every common word any player would want, plus enough depth for interesting discoveries. Different applications might threshold differently—a crossword game might want terms at rank 200K that a casual word game would skip.

Definitions

Linguabase definitions are readable paragraphs—2–3 sentence blurbs that cover all a word’s meanings naturally, not dictionary-style numbered fragments.

Why LLMs Can Do This (Sort Of)

This is the layer where LLMs come closest to being a substitute. You could brute-force 400K API calls to generate definitions in your style. But there are catches:

• You need a word list first—LLMs don’t know which 400K words you want defined.
• Capitalization confusion—LLMs routinely conflate “china” (porcelain) with “China” (country), “march” (walk) with “March” (month). This contaminates thousands of words unless you handle it carefully.
• Sense imbalance—LLMs often have trouble covering all the senses of a word in a balanced way.
• Style consistency—Maintaining a consistent voice across 400K definitions is hard. Our definitions were written to a unified editorial standard.
• Not all studios can manage bulk API work—400K API calls, error handling, rate limits, cost management. It’s doable, but it’s engineering overhead you may not want.

Content Filters

For applications with content policies, Linguabase provides two word lists you control:

Hard Block List

A few thousand words including vulgar terms and creative permutations of sexual and racial slurs. These words exist in dictionaries but often have no legitimate use in puzzle-making or public/sharing contexts.

Soft Block List

Many words are technically inoffensive but may be inappropriate for automated puzzle generation:

• Contaminated by sound—“niggardly” (meaning stingy, from Old Norse) is unrelated to the racial slur, but its phonetic similarity makes it unusable in modern contexts
• Multiple senses—“knockers” (door hardware vs. slang), “stroke” (medical/sports vs. suggestive), “spank” (innocent action vs. suggestive)
• Context-dependent—“moist” (technically inoffensive but widely disliked), “threesome” may be fine in numerical contexts but problematic in games

We developed the soft-block list to exclude these words from automated puzzle generation. Though these words should typically be allowed as valid user-entered answers. You might not want to generate a puzzle featuring “knockers” as a target word, but it could be a perfectly acceptable player-provided answer in a spelling game about doors.

Word Associations

Words relate to other words in different ways. Linguabase handles three distinct types of multi-meaning relationships:

Each type requires different handling. Homographs need complete separation—eye-pupil associations shouldn’t contaminate student-pupil associations. Polysemous words need their branching senses tracked. Facets can be blended or kept separate depending on your use case.

Core Associations

The core associations give you approximately 40 related words per entry, weighted by relationship strength. This represents an amalgam of different meanings, connotations, and facets of a word:

Drilling Deeper into Meaning

Beyond the core associations, Linguabase provides pools of words that reflect narrow, particular meanings of any headword:

Why Not Just Use an LLM?

You could prompt an LLM for word associations. Superficially, it looks fine. But problems lurk under the surface:

• Flat experience in aggregate—LLMs cluster on expected associations from training data. Ask for 1,000 different words and you’ll get similar patterns: the dominant sense, the obvious connections. Your users will notice the sameness over time.
• Sense imbalance—LLMs often have trouble covering all senses of a word in a balanced way. “Elephant” has 9 distinct senses in Linguabase (anatomy, behavior, circus, heraldry, ivory, megafauna, size, symbolism, wildlife)—LLMs typically cluster on just 2-3.
• Edge cases everywhere—Capitalization (polish vs Polish), false cognates (dig → digress), superconnector words that dominate everything, function words with no associations. Each requires special handling.

LLM inferences are a major input signal that we use, with multiple query types feeding into our ranking and validation pipeline. But in our experience, LLMs serve as better editors than authors. Our workflow pools data about a word to ask rich questions to an LLM—like “which of the following words are strongly related?”—instead of asking the LLM to provide all the answers. More on LLM limitations →

False Cognate Removal

False cognates are words that look related but aren’t—they share spelling patterns by coincidence, not common origin. String similarity filters can’t detect them, and LLMs sometimes hallucinate their associativeness:

We used LLM-based auditing to identify false cognates across all headwords.

Capitalization Intelligence

Since lowercase words are capitalized at the beginning of sentences, LLMs often treat text case-insensitively, conflating “polish” (verb) with “Polish” (nationality). For semantic graphs, this contaminates association lists—“Poland” shouldn’t appear in the associations for shoe polish.

We evaluated capitalization variants for ambiguous terms, testing whether each form works in natural sentences. Results:

• 3,509 words have two valid capitalization variants (like turkey/Turkey)
• 86 words have three variants (like swat/Swat/SWAT)
• Each variant gets its own decontaminated association list

Common Word Coverage

Traditional thesauruses treat function words as “stopwords”—ignored entirely. Linguabase does the opposite: we put extra effort into building associations for hundreds of the most common words that other sources systematically skip.

These words are the glue of language. When your game needs to understand how “and” connects to other concepts, we have answers that a thesaurus typically won’t provide.

Gestalt Enrichment

One source of richness in Linguabase is gestalt relations—experiential and sensory associations that taxonomic approaches miss entirely. These are folded into the core associations, enriching what you get for each word.

These are NOT synonyms. They’re how humans actually experience concepts. A thesaurus typically won’t tell you that “crisis” evokes “siren”—but humans know this instantly.

Word Families

Morphological and etymological groupings that connect related word forms:

Variant vs Semantic Distinction

Word families serve two different purposes, depending on your game design needs:

Why this matters:

• “butterfly → butterflies” = variant swap (same concept, just plural)
• “butterfly → fly” = semantic hop (different concept via shared morpheme)

But not everything that looks splittable should be split. “Mushroom” contains “room” and “mush,” but it’s not a compound of those words—it comes from French mousseron. String patterns alone can’t distinguish real morphological relationships from coincidental letter sequences. This is why the etymological audit matters.

Usage Examples

1.46M illustrative quotations written by humans with intent. Almost all common words come from famous literature—Pulitzer Prize winners and other prestigious sources. Uncommon and technical words are sourced from Wikipedia or open-access science journals, still showing the word used in real context by real writers.

The Data Pipeline

Linguabase is built on a decade of work that began before LLMs existed. The foundation is human curation and pre-LLM computational linguistics; modern LLMs handle validation at scale.

Three Phases: Expand, Audit, Contract

Our pipeline follows a consistent pattern for each data layer:

The result: we start with millions of candidate relationships and ship ~40M high-quality connections.

Development Timeline

Professional Lexicography

Orin Hargraves (professional lexicographer, contributor to major dictionaries) wrote 2,000+ custom definitions and 4,400+ sense-grouped word associations. His work focused on the words that need the most human judgment: interjections, prepositions, the highest-frequency words, and words with so many dictionary senses that their entries are virtually unreadable.

Linguistics grad students and post-docs created 5,000+ thematic word lists over several years:

• Parts of a watch, types of gargoyles, painting supplies
• Cross-domain connections and specialized vocabulary
• Edge cases that automated approaches systematically miss

Reference Sources

We integrated 70+ linguistic resources—professional thesauri like the NASA Thesaurus, public domain lexicons like WordNet and Roget’s (we explain why these aren’t enough on their own), Library of Congress subject headings, and specialized vocabularies spanning scientific, governmental, artistic, and medical domains. Each source required custom parsing and integration; the specific methods reflect years of experimentation about what works for word games versus dictionary lookup.

Pre-LLM Computational Linguistics

Before LLMs existed, we built the initial graph using classical NLP:

• Wikipedia extraction—Link analysis, disambiguation pages, definition summaries
• Latent Dirichlet Allocation—2.3M supercomputer hours (NSF XSEDE grant, 2013-14) for topic clustering across massive corpora
• Word2Vec models—Statistical word embeddings trained on billions of tokens

These methods produced the raw material—millions of candidate relationships—that human curation and later LLM validation refined into production-quality data.

Library of Congress Expansion

We methodically processed all 648,000 Library of Congress subject classifications—capturing the themes of millions of books humans wanted to write in English. Human librarian classifiers organized these into topics like “orange horticulture” or “Indus Valley civilization.” By analyzing pools of words across these classifications, we discovered semantic clusters that no dictionary or thesaurus would surface. These serve as “idea seeds” for expanding our association graph into domains that traditional lexicography misses.

LLM Validation

We use LLMs for validation, not generation. The difference matters—see why LLMs can’t generate this data reliably. an LLM can confirm that “key → reef” is a valid association (it’ll say yes), but it won’t generate that association reliably on its own. Our pipeline proposes candidates from the sources above, then uses LLM scoring to rank, filter, and audit them.

Production Refinement

Linguabase has powered “In Other Words” since 2011. Countless edge cases have emerged through real gameplay that no automated process would surface:

• Words that look related but aren’t (false cognates discovered through player confusion)
• Cultural associations that shift over time
• Difficulty calibration based on actual player success rates
• Content filter refinements from user reports

Ongoing Curation

Beyond automated processing, our manual override layer includes 50,000+ hand-curated entries (definitions, sense-grouped associations, thematic lists) plus hundreds of thousands of corrections for plurals, word families, and edge cases.

This is the work that separates production-quality data from raw output—and it’s ongoing. Each pipeline run incorporates feedback from the previous one.

Why 1.5M Internally, 400K Deployed?

Internally, we maintain 1.5 million words—including all of Wiktionary plus the top 200,000 words from Wikipedia. But we don’t ship all of these because noise gets in. We don’t include “Declaration of Independence” even though it’s in our word list; we don’t include obscure proteins that no player would recognize. The 400K threshold captures every word players actually want, without the noise. The hard part isn’t building a huge list—it’s ranking and curating it.

Game Mechanics Based on Word Relationships

Because the data is structured as a weighted graph, you can build game mechanics based on associations and linkages. The weighted connections enable real-time content customization—filtering by relationship strength, guiding players toward goals, or generating puzzles with tunable difficulty. (Or skip the engineering—we can generate puzzle data for your game mechanics directly.)

Pathfinding

Find how any two words connect through the graph:

We generate pathfinding puzzles like this at scale—with validated paths, precalculated hints, and difficulty ratings.

Convergence

If you’re building a game that involves non-adjacent word associations (like a chain-association game), real-time brute-force graph analysis has exponential expansion that cannot be completed on modern hardware. But using our data, you can pre-calculate convergence probabilities to targets.

Given a target word, which intermediate words get you closer? Consider “ellipse”—its associations organize into layers by distance:

Paths using Layer 1 terms are easier; paths requiring Layer 2–3 terms are progressively harder. This enables hint systems and adaptive difficulty. See Puzzle Licensing for how we package this into game-ready data.

Small World Property

Analysis of the Linguabase graph shows that 76% of English word pairs connect in 7 hops or fewer. Average path length is 6.43 steps.

The hop-distance distribution peaks at 5-6 hops:

Of the 1.5M headwords, about 870K (57%) are reachable through the top-40 associations of other words. The remaining 43% are rare or isolated terms that don’t appear in other words’ association lists.

This is important because it’s a foundational property of the English language that you can exploit to create your own games. The semantic space is more connected than people realize.

Continuous Improvement

Linguabase is actively maintained, not a static dataset:

• New vocabulary—Words enter mainstream usage constantly (“doomscrolling,” “finsta,” “situationship”). We add them as they stabilize.
• Quality refinement—Each pipeline run improves grades based on new validation data
• Sense updates—Meanings shift over time; we track semantic drift
• False cognate auditing—Ongoing review catches edge cases

License terms include current data with update arrangements available.