WordNet Expansion

The core mechanism of WordNet Expansion relies on synsets—sets of cognitive synonyms that represent a single lexical concept. When a query term is identified, the system retrieves its corresponding synset and adds all member words to the query.

• Example: A query for 'automobile' expands to include 'car', 'auto', 'machine', and 'motorcar'
• Each synset is assigned a unique offset ID for precise database lookups
• Disambiguation is critical: 'bank' as a financial institution vs. 'bank' as a river edge belong to entirely different synsets

WordNet organizes nouns into a lexical inheritance hierarchy using hypernym (is-a) relationships. Expansion can traverse upward to broaden queries or downward to specialize them.

• Hypernym Expansion: A query for 'terrier' broadens to include 'dog', 'canine', 'carnivore', and 'animal'
• Hyponym Expansion: A query for 'furniture' narrows to include 'chair', 'table', 'sofa', and 'bed'
• The hierarchy forms a directed acyclic graph, not a strict tree, allowing multiple inheritance paths

Beyond class hierarchies, WordNet captures part-whole relationships that enable compositional query expansion. Meronyms name constituent parts, while holonyms name the containing whole.

• Part Meronym: 'finger' is a part of 'hand'; 'hand' is a part of 'arm'
• Member Meronym: 'player' is a member of 'team'; 'tree' is a member of 'forest'
• Substance Meronym: 'wood' is the substance of 'table'; 'flour' is the substance of 'bread'
• These relations are especially valuable for e-commerce and technical documentation search

WordNet partitions its database by part of speech—nouns, verbs, adjectives, and adverbs—each with distinct relational structures. Query expansion must respect these boundaries to avoid nonsensical term additions.

• Nouns: Organized primarily by hypernym hierarchy
• Verbs: Structured by troponymy (manner-of relations); 'walk' → 'stroll', 'march', 'saunter'
• Adjectives: Organized by antonymy and similarity clusters; 'hot' ↔ 'cold'
• Adverbs: Derived from adjectives with limited relational structure
• A query term tagged as a noun will not expand using verb synsets, preventing semantic drift

WordNet 3.1 includes domain labels and usage notes that constrain expansion to contextually appropriate terms. This prevents archaic, slang, or domain-specific terms from polluting general-purpose queries.

• Register labels: 'slang', 'formal', 'colloquial' filter inappropriate synonyms
• Domain categories: Terms tagged with 'medicine', 'law', 'computing' can be selectively included or excluded
• Frequency counts: Each sense is ranked by corpus frequency, allowing expansion to prioritize common usages
• Example: The medical sense of 'virus' expands differently than the computing sense

Before WordNet lookup, query terms must undergo lemmatization to map inflected forms to their base lemma. WordNet stores only canonical forms, so 'running', 'ran', and 'runs' must all reduce to 'run'.

• Morphy: WordNet's built-in morphological analyzer handles regular and irregular inflections
• Exception lists: Hardcoded mappings for irregular forms like 'children' → 'child', 'better' → 'good'
• Compound handling: Multi-word expressions like 'hot dog' or 'kick the bucket' are stored as collocations
• Failure to lemmatize correctly is the most common source of expansion errors in production systems

Adds terms with identical or highly similar meanings to the original query to increase recall without altering scope.

• Replaces 'automobile' with 'car'
• Often the first expansion step before hierarchical traversal
• Can be derived from WordNet synsets or distributional thesauri

Key distinction: Synonym expansion stays at the same level of specificity, unlike hypernym/hyponym expansion which moves vertically in the hierarchy.

Broadens a query by adding more general terms from the semantic hierarchy.

• Adds 'vehicle' to a query for 'car'
• Increases recall at the cost of precision
• Critical for bridging the vocabulary gap when users are too specific

Risk: Over-generalization can introduce noise. Typically weighted lower than exact matches in retrieval scoring.

Narrows or specifies a query by adding more specific subordinate terms.

• Adds 'sedan', 'SUV', and 'coupe' to a query for 'car'
• Increases precision for ambiguous queries
• Useful when the user's term is too broad for the target corpus

Implementation: Requires traversing the WordNet hyponym tree with a configurable depth parameter to avoid over-specification.

Assumes the top-k documents from an initial retrieval are relevant and extracts expansion terms from them automatically.

• No external lexical database required
• Adapts to the corpus's specific vocabulary
• Computationally expensive due to two-pass retrieval

Contrast with WordNet: PRF is corpus-driven and dynamic; WordNet expansion is knowledge-driven and static. PRF can capture domain-specific jargon that WordNet misses.

Uses deep language models like BERT to generate expansion terms that are semantically appropriate for the query's specific context.

• Disambiguates polysemous words (e.g., 'bank' as river vs. financial)
• Captures semantic similarity beyond rigid lexical relationships
• Requires significant computational resources at query time

Advantage over WordNet: Handles domain-specific and figurative language that falls outside structured lexical databases.

Enriches a query by traversing an entity's relationships in a knowledge graph to add related entities and attributes.

• Adds 'Elon Musk' and 'Tesla' to a query about 'SpaceX'
• Leverages typed edges (founderOf, competitorOf) for precise expansion
• Requires a well-maintained, domain-specific knowledge graph

Synergy with WordNet: Knowledge graphs handle named entities and factual relationships; WordNet handles common nouns and lexical semantics. Production systems often combine both.