Mapping Language: The IDEA Linguabase

The IDEA Linguabase combines traditional lexicography with modern data processing and large language models. With 1.1 million headwords connected by 60 million weighted relationships, this language database extends well beyond traditional reference works. This unprecedented scale reflects our inclusion of everyday objects, multi-word phrases, and encyclopedic terms—not just the abstract concepts found in standard thesauri.

Beyond Traditional Thesauri

Traditional thesauri serve as “synonym dictionaries” – references where writers find different words with similar meanings. These works typically focus on abstract concepts, emotions, actions, and qualities rather than concrete objects, for practical reasons:

1. Writers need synonyms for verbs, adjectives, and abstract nouns more than for concrete objects
2. Abstract terms like “applause” connect naturally to many related concepts (acclaim, ovation, praise), while concrete nouns like “apple” have fewer true synonyms
3. Physical thesauri faced space constraints, forcing editors to prioritize frequently needed alternatives

Linguabase breaks from this tradition in several ways:

• Comprehensive coverage: Unlike traditional thesauri that might omit everyday objects (containing “applause” and “appliance” but not “apple” or “apple pie”), Linguabase includes all words, including concrete nouns, specialized terminology, common objects, and thousands of encyclopedic proper nouns
• Multiple relationship types: Beyond synonyms and antonyms, Linguabase maps:
  • Similar meanings: Words with close semantic relationships like “house,” “domicile,” and “lodge”
  • Category members: Items of the same type such as “house,” “bungalow,” and “villa”
  • Associative relationships: Words contextually related like “house,” “quarter,” and “dwell”
• Weighted connections: Each relationship carries a decimal score (scores above 1 indicate strong correlation; scores between 0 and 1 represent lower confidence associations)

This approach provides an average of 60 semantically connected words for each headword across all parts of speech, covering multiple senses and contextual usages that traditional reference works typically omit.

Comprehensive Linguistic Coverage

Linguabase distinguishes itself through four key advantages:

• Unparalleled scale: 1.1 million headwords far exceed traditional lexical databases like Princeton’s WordNet (1985)
• Extensive relationship network: 60 million weighted connections provide structure that large language models lack
• Multiple relationship types: Coverage extends beyond synonyms to include categorical, contextual, and associative connections
• Human oversight with AI enhancement: Human-curated content augmented by artificial intelligence

Multi-Sense Representation

A crucial feature of Linguabase is its handling of words with multiple meanings:

• Double meanings (technically called “homographs”): Words spelled identically but with entirely different meanings, often with different origins and pronunciations. English contains approximately 1,000-3,000 of these.
  • Example: “bass” (low sound) vs. “bass” (type of fish)
  • Example: “tear” (drop from eye) vs. “tear” (to rip)
• Related meanings (technically called “polysemes”): Words with multiple distinct but connected definitions that have evolved from the same root, typically appearing as separate numbered entries in dictionaries
  • Example: “head” (body part, leader of organization, front of ship)
  • Example: “branch” (tree limb, division of organization)
  • Example: “hiking” (walking on trails for recreation) vs. “hiking” (forcefully moving something upward, as in “hiking prices”)
• Contextual flavors: Different aspects or dimensions of the same meaning that emphasize different connotations depending on context
  • Example: “hiking” as recreation can emphasize either nature aspects (outdoors, scenery, wildlife) or exercise aspects (exertion, fitness, calorie-burning)
  • Example: “coffee” as a beverage vs. a social ritual
  • Example: “reading” as education vs. entertainment

Unlike polysemes which have distinct definitions, contextual flavors describe how words activate different associative networks while retaining the same core meaning. This distinction captures how people actually use and understand language in everyday contexts.

Building the Database: Four Knowledge Sources

The creation of Linguabase involved four complementary knowledge sources that feed into an amalgamated scoring system:

1. Reference Integration

We analyzed over 70 distinct lexicographic resources, including: Wiktionary, WordNet, Getty Art & Architecture Thesaurus, AGROVOC Thesaurus, Library of Congress Subject Headings, NASA Thesaurus, National Library of Medicine’s UMLS Metathesaurus, USDA National Agricultural Library Thesaurus, Moby Thesaurus II, Roget’s Thesaurus variants, and the Ethnographic Thesaurus.

This integration process combined relationships from multiple sources, with repeated occurrences of a relationship across multiple sources naturally boosting its confidence weight.

2. Topic Modeling

To capture broader associations between words, we applied topic modeling to extensive collections of English prose. For each term, we extracted matching sentences and paragraphs, then used Latent Dirichlet Allocation—a statistical method that discovers abstract topics as collections of words that frequently appear together—to identify approximately 8 abstract topics per analysis. This computation-intensive analysis required supercomputing resources from the NSF-funded Extreme Science and Engineering Discovery Environment (XSEDE). The distributed processing allowed us to analyze thousands of text samples and generate millions of weighted word relationships in days rather than years.

3. Structured Word Groups

Linguistics graduate students created over 10,000 word groups based on Library of Congress categories. These categories are significant because they were designed to organize millions of books written by countless authors throughout history, inherently reflecting the vast range of topics that writers have ever wanted to discuss.

The Library of Congress classification system uses a hierarchical structure, beginning with broad parent categories (examples below) that branch into thousands of highly specific subcategories:

• AC: Collections, Series, Collected works
• BF: Psychology, Parapsychology, Occult sciences
• GN: Anthropology, Ethnology, Folklore
• HQ: Family, Marriage, Women, Sexuality
• KF: United States federal law

For example, under “QK” (Botany), our word groups included specific concepts like QK495 (Classification of plants as angiosperms), QK917 (Plant ecology and carnivorous plants). Similarly, within “VM” (Naval architecture), subcategories like VM156 (Shipbuilding materials), VM311 (Hull design), and VM747 (Marine engines) each generated multiple specialized vocabulary sets.

This approach provided domain-specific vocabulary coverage across all fields of knowledge, including specialized terminology and high-frequency terms typically omitted from reference works.

4. Large Language Model Enhancement

Linguabase uses advanced large language models to supplement the existing structured data in several crucial ways:

• Expanding coverage for everyday terms: These models excel at generating rich associations for common terms like “apple pie” that lack substantial coverage in traditional reference works but evoke strong connections for most speakers
• Handling morphological variations: While traditional thesauri might contain “apply” but not “applies,” language models help generate complete paradigms across different parts of speech and inflected forms
• Identifying contextual flavors: Modern language models proved essential for recognizing and mapping the subtle connotative dimensions of words in different contexts
• Managing capitalization distinctions: These systems effectively differentiate between terms like “China” (country) vs. “china” (porcelain) or “Trump” (surname) vs. “trump” (card game), and correctly capitalize terms within the data graph
• Processing compound words: Advanced models naturally handle multi-word expressions like “New York” or “department store” without the parsing difficulties that traditional computational approaches often encounter
• Optimizing relationship rankings: In applications where only a few “best” word relations can be displayed, language models improved the prioritization of the most relevant connections

This strategic use of large language models enhances areas where traditional lexicographic approaches have limitations while preserving the reliability of human-curated content for core semantic relationships.

Data Processing and Weights

Linguabase employs a practical approach to relationship weighting. The 60 million weighted relationships derive from:

• Frequency of appearance across multiple sources
• Editorial judgment about relevance and association strength
• Statistical significance from topic modeling
• Semantic proximity determined by language models

The database operates through batch processing with targeted updates. A full rebuild takes approximately one week on consumer hardware. This approach prioritizes useful results over methodological complexity.

Practical Applications

Linguabase powers two applications:

In Other Words: Word Exploration Game

This interactive game allows players to navigate between concepts using meaningful connections. Players can find paths between seemingly unrelated words by traversing the weighted relationship network, demonstrating both the breadth and interconnectedness of language.

Comprehensive Reference System

Our reference application extends traditional thesaurus functionality by:

• Providing relationships for concrete objects traditionally omitted from thesauri
• Showing multiple senses and contextual variations for each word
• Offering weighted connections that indicate relationship strength
• Including associative relationships beyond simple synonyms

Practical Innovation

The value of Linguabase lies in its thorough integration of classical lexicography with modern language models. By combining multiple knowledge sources and enhancing them with advanced AI, we’ve created a resource that’s:

• More comprehensive than traditional lexical databases
• More structured than raw large language model output
• Accessible for both specialized NLP applications and everyday language exploration

This approach demonstrates how traditional linguistic resources can be augmented rather than replaced by large language models, creating a practical tool that advances the field of lexicography.

The IDEA Linguabase was developed with support from the NSF-funded Extreme Science and Engineering Discovery Environment (XSEDE), grant #IRI130011.