Knowledge Graph Expansion

The core mechanism involves identifying a seed entity in the query and then walking the graph's edges to discover connected nodes. Unlike synonym expansion, which relies on lexical similarity, this technique uses semantic relationships like isA, hasProperty, or locatedIn.

• 1-hop expansion: Adds directly connected entities (e.g., for 'Tesla', add 'Elon Musk' via foundedBy).
• 2-hop expansion: Traverses further to add context (e.g., 'Elon Musk' -> 'SpaceX' via CEOof).
• Path ranking: Algorithms score the relevance of a traversal path to avoid introducing noise from overly generic connections.

Beyond adding related entities, expansion enriches the query with specific literal properties stored in the graph. This is critical for structured search over unstructured text.

• If a query mentions 'iPhone', the graph can inject attributes like manufacturer: Apple or releaseYear: 2023.
• This converts a vague keyword search into a faceted, structured query without requiring the user to know the exact schema.
• SPARQL or Cypher queries are dynamically generated to pull these attributes based on the entity's type in the ontology.

Effective expansion requires resolving the query entity to a single, unambiguous node in the graph before traversal begins. Expanding from the wrong entity (e.g., 'Mercury' the planet vs. the element) catastrophically degrades precision.

• Entity linking is a prerequisite step, using context words in the query to score candidate nodes.
• Popularity priors and type checking (e.g., the query context implies a Person type) are used to break ties.
• The system must often decide to not expand if the disambiguation confidence is below a strict threshold.

To meet the sub-100ms latency requirements of search, knowledge graphs used for expansion are heavily optimized for read speed. This involves pre-computing traversal paths and denormalizing data.

• Adjacency matrices and lookup tables replace real-time deep graph queries.
• Frequently accessed 1-hop neighborhoods are cached in memory.
• Graph embeddings (e.g., TransE, RotatE) can be used to pre-compute entity similarity, allowing for a vector-based nearest-neighbor search over the graph structure itself, bypassing explicit edge traversal.

Knowledge graph expansion provides deterministic, symbolic reasoning, which contrasts sharply with the statistical similarity of embedding-based expansion.

• Graph expansion: Adds 'Paris' to 'France' via the capitalOf relationship. This is a logical fact.
• Embedding expansion: Adds 'Paris' to 'France' because their vectors are close in training data. It might also erroneously add 'Lyon' or 'Germany'.
• Graph expansion is preferred in high-precision domains like legal or medical search where hallucinated relationships from embeddings are unacceptable.

Expanded entities and attributes are not treated equally with the original query terms. They are assigned down-weighted boost factors to ensure the original user intent remains primary.

• A common strategy is a decay function: a 1-hop expansion term gets a boost of 0.5, a 2-hop term gets 0.25.
• Inverse entity frequency: Rare, highly specific entities from the graph get a higher boost than common, generic ones.
• This weighting is applied at the retrieval stage, often as a should clause in a boolean query with a boost parameter.