Detects entities in user queries and rewrites the query to match how documents naturally reference those entities, recovering relevant documents that would not match the literal query but do match the entity-aware rewrite.
Patent Overview
- Inventor
- Krishna Bharat
- Assignee
- Google LLC
- Filed
- 2004-04-06
- Granted
- 2009-05-19
- Application Number
- US 10/818,540
The Challenge
The Challenge
Users phrase queries one way; documents reference the same entities another way. 'Obama's birthplace' and 'Where was Barack Obama born' should retrieve the same documents but literal-match retrieval misses the connection. Entity-aware rewriting bridges the gap.
- Surface Phrasings Differ From Document Phrasings — Users abbreviate, paraphrase, and use colloquial entity names. Documents use canonical names, full titles, and structured references. Literal match misses the bridge.
- Synonym Expansion Is Too Broad — Naive synonym expansion adds many irrelevant terms and dilutes precision. Entity-aware expansion targets specific entities cleanly.
- Entity Identification Must Be Reliable — Wrong entity identification rewrites the query to mean something else. The detector must achieve high precision before the rewriter can act on it.
- Rewrites Must Preserve User Intent — An aggressive rewrite that changes the query's meaning hurts more than helps. Rewrites must add entity-aware alternatives, not replace the original query intent.
- Cross-Language Variants Matter — Multi-lingual entity names should map to a single canonical entity. The system needs entity normalization across languages and writing systems.
Innovation
How The System Works
The system detects entities in the query using an entity-recognition layer, retrieves alternative canonical references for each detected entity, generates rewritten query variants, retrieves documents against both the original and rewritten queries, and merges results so the user sees an expanded but precise candidate set.
- Identify Entities In The Query — Entity recognizer scans the query for known entities. Detected entities resolve to canonical IDs in the entity graph.
- Retrieve Alternative References — For each entity, retrieve canonical alternatives: full official name, common abbreviations, well-known synonyms. The set captures how documents actually reference the entity.
- Generate Rewritten Variants — Combine the original query structure with alternative entity references to generate query variants. Each variant preserves the query intent while substituting entity phrasing.
- Score Variants For Quality — Each variant is scored on grammatical naturalness, retrieval likelihood, and intent preservation. Low-quality variants are dropped.
- Run Multi-Variant Retrieval — Each surviving variant retrieves its own candidate set. The original query also retrieves.
- Merge And Deduplicate Results — Candidate sets merge with deduplication. The merged set is the expanded candidate pool.
- Rank Merged Pool — The merged candidates go through standard ranking. The user sees results that include both literal-match documents and entity-aware-rewrite documents.
Entity-Aware Query Expansion
The patent's load-bearing idea is to use entity recognition as the bridge between user phrasing and document phrasing. Synonym expansion is too broad; entity-aware expansion targets specifically the entities that need bridging.
Bridge The Phrasing Gap Per Entity
Different entity references appear in user queries vs documents. Detecting the entities and substituting their canonical references closes the gap precisely, without the noise of broad synonym expansion.
- Entity Recognition — Reliable entity recognition is the prerequisite. Wrong entities produce wrong rewrites; only high-precision detection unlocks the value.
- Canonical Reference Lookup — Each detected entity links to a set of canonical alternatives: official name, common abbreviations, synonyms. The alternatives drive rewriting.
- Multi-Variant Retrieval — Each rewritten variant retrieves its own candidates. Merging produces an expanded pool that captures both literal and entity-aware matches.
Technical Foundation
Technical Foundation
The patent specifies the entity recognizer, the alternative reference store, the rewrite generator, the variant scorer, and the multi-retrieval merger.
- Entity Recognizer — Neural model identifies entities in the query and resolves them to canonical IDs. Trained on entity-tagged query corpora plus the entity graph.
- Alternative Reference Store — Per-entity, the store holds canonical alternatives: official name, abbreviations, well-known aliases, cross-language variants. The store is sourced from the entity graph plus query log mining.
- Rewrite Generator — Template-based and neural rewriters produce query variants combining original structure with alternative entity references. The output is a set of candidate variants.
- Variant Scorer — Each variant is scored on grammatical naturalness, query-log frequency (whether the variant is a real query users issue), and intent preservation.
- Multi-Retrieval Engine — The retrieval engine accepts multiple queries simultaneously and returns their results. Latency is bounded by the slowest variant.
- Result Merger — The merger deduplicates results across variants and produces the unified candidate set for downstream ranking.
The Process
The Process
The rewriting pipeline runs in the query path. Latency overhead is small because entity recognition and rewriting are fast; the main cost is the multi-variant retrieval.
- Query Enters Pipeline — User query arrives at the query parser. The parser hands it to the entity recognizer.
- Detect Entities — Entities are detected and resolved to canonical IDs. Entity-free queries skip the rewriting path.
- Look Up Alternatives — For each entity, alternatives are retrieved from the reference store. The set is small and bounded.
- Generate Variants — The rewriter produces query variants by substituting alternative references into the original structure. Variants preserve intent.
- Score And Filter Variants — Low-quality variants are filtered. The output is a small high-quality variant set.
- Multi-Retrieval Execution — The original plus each variant retrieves its candidates in parallel. Results stream back to the merger.
- Merge, Rank, Return — Merged candidates go through standard ranking. The user sees results from both literal and rewrite paths.
Quality Control
Quality Control
Wrong rewrites degrade precision. The patent specifies safeguards against entity-recognition errors and variant-quality drift.
- Entity Recognition Confidence Threshold — Only high-confidence entity detections trigger rewriting. Low-confidence detections fall back to literal-match retrieval.
- Variant Quality Floor — Variants below the quality floor are dropped. The system prefers fewer high-quality variants to many low-quality ones.
- Intent Preservation Verification — Each variant is checked to ensure it preserves the original query's intent. Variants that drift in meaning are excluded.
- Result Overlap Audit — If a variant retrieves a substantially disjoint result set from the original, the variant is flagged for inspection. Disjoint results usually indicate intent drift.
- Alternative Reference Curation — The reference store is curated. Wrong or stale alternatives are removed; new ones added as entities accumulate canonical references.
Real-World Application
Query rewriting with entity detection underpins Google's entity-aware retrieval, Knowledge Panel triggering, and the natural-language query handling in modern Search. The primitives generalize to voice queries and conversational search.
- Entity-targeted Rewrite Scope — Only entities trigger rewriting. The system avoids the noise of broad synonym expansion by targeting specific entity references.
- Multi-variant Retrieval Strategy — Each rewritten variant retrieves its own candidates. Merging produces an expanded pool that catches both literal and entity-aware matches.
- Confidence-gated Safety Mechanism — Low-confidence entity detection falls back to literal-match. Wrong rewrites are avoided rather than risking precision.
Why Canonical Entity Naming Compounds Visibility
Content using canonical entity names matches retrieval whether the user queries with canonical or alternative phrasing. The rewriter bridges the gap to canonical content; non-canonical content has to hope the rewriter happens to substitute its specific phrasing.
Why Schema Markup Helps Entity-Aware Retrieval
Pages with explicit entity markup help the rewriter know which entities a page is about. The page is then a high-confidence retrieval target for entity-aware rewrites, increasing visibility on entity-bearing queries.
<\/section>What This Means for SEO
What This Means for SEO
The patent detects entities in queries and rewrites to their canonical references, retrieving documents that match the entity-aware rewrite even when they miss the literal query. SEO implication: using canonical entity names matches retrieval regardless of how the user phrases the entity, so canonical naming compounds visibility.
- Canonical Entity Naming Compounds Visibility — Content using canonical entity names matches retrieval whether the user queries with canonical or alternative phrasing, because the rewriter bridges to canonical content. Non-canonical content has to hope the rewriter happens to substitute its specific phrasing.
- Schema Markup Helps Entity-Aware Retrieval — Pages with explicit entity markup help the rewriter know which entities a page is about, making it a high-confidence retrieval target for entity-aware rewrites. Markup increases your visibility on entity-bearing queries.
- Entity Recognition Bridges Phrasing Gaps — The patent bridges user phrasing and document phrasing per entity, more precisely than broad synonym expansion. Naming entities the way documents canonically reference them puts you on the matched side of the bridge.
- Bridge Targets Specific Entities — Entity-aware expansion targets the detected entities, not the whole query. Clearly establishing which entities your page covers makes it a precise target for rewrites of queries about those entities.
- Alternative Phrasings Route To Canonical Content — Users querying with informal or alternative names get routed to canonical content via rewriting. Owning the canonical reference for an entity means you capture the alternative-phrasing queries that get rewritten toward it.
- Consistency Aids Detection — The rewriter relies on detecting entities reliably. Referring to entities consistently and unambiguously across your content helps detection identify your page as about those entities, improving rewrite-driven retrieval.
- Merged Results Reward Both Phrasings — The system merges results from original and rewritten queries. Content covering both the canonical name and common alternatives can match on both retrieval paths, maximizing presence in the merged candidate set.