Combines a query-document relevance score with a similarity score derived from a feature vector of document attributes and query-word presence, producing rank values that learn from query context.
Patent Overview
- Inventor
- Prabhakar Raghavan
- Assignee
- Verity, Inc.
- Filed
- 2001-05-08
- Granted
- 2004-05-18
- Application Number
- US 09/851,675
The Challenge
Static Ranking Functions Miss Contextual Relevance
Classical search ranking uses a single relevance score derived from term frequency, inverse document frequency, and link authority. This single-score model is blind to query-document context. Two documents with the same relevance score may serve the query very differently in practice. The system needs a ranking method that adapts to query context by reading attributes of the document and how they align with the specific query.
- Single Score Misses Context — A relevance score alone cannot distinguish between two documents that score identically on TF-IDF but address the query differently. Context is invisible to single-score ranking.
- Documents Have Attributes Beyond Text — Documents carry attributes (length, format, freshness, structure) that matter for ranking but are not captured in classic relevance models. The ranking needs to consume these explicitly.
- Query-Word Pattern Is A Signal — How query words distribute across a document (positions, density, structural placement) reveals more than raw term frequency. The pattern should be a ranking input.
- Adaptivity Needs A Combined Score — Adapting to query context requires combining multiple scoring axes. A single number is insufficient; a feature vector plus a learned combiner is needed.
- Feature Vector Comparison Is The Mechanism — Comparing a query feature vector to a document feature vector via similarity is what allows context-aware ranking. The similarity captures alignment that pure relevance does not.
Innovation
Relevance Score Plus Similarity Score
Each document gets two scores. A relevance score reflects standard query-document term match. A similarity score is computed by comparing a feature vector that characterizes both the document's attributes and the query words associated with it. The rank value combines both, producing context-aware ranking that adapts as the system learns which features matter for different query types.
- Produce Relevance Score — Compute the standard relevance score for the (document, query) pair using TF-IDF, BM25, or another classical relevance function.
- Construct Document Feature Vector — Build a feature vector that captures document attributes and how query words appear in the document: positions of query words, density distribution, structural placement, document length, format.
- Construct Query Feature Vector — Build a complementary feature vector for the query that captures its attributes: length, query-class signals, structural form.
- Compute Similarity Score — Calculate similarity between the document feature vector and the query feature vector. Cosine similarity is the typical choice; other vector similarity measures work too.
- Combine Into Rank Value — Combine the relevance score and the similarity score into a single rank value. Weights are tunable and can be learned from training data with relevance judgments.
- Order Results By Rank Value — Sort the candidate documents by the combined rank value. The rank value is what drives ordering in the final result list.
Two Scoring Axes, Adaptive Ranking
Treating ranking as a combination of relevance and similarity-of-features is the patent's structural contribution. The system can adapt by tuning the combination weights without changing either underlying score.
Relevance Tells Topic; Similarity Tells Context
The relevance score says the document is about the query topic. The similarity score says the document is structured the way users expect for this query type. Both are needed.
- Relevance Score — Standard query-document term-match measure. Necessary for being on-topic.
- Similarity Score — Feature-vector alignment between document attributes and query attributes. Captures structural and contextual fit.
Ranking becomes adaptive by combining scores that respond to different signals.
<\/section>Technical Foundation
What The Feature Vector Captures
The feature vector is the central novel piece. Different choices of features produce different adaptive ranking behaviors.
- Document Attributes — Static properties of the document: length, format, freshness, structural complexity. Available offline at index time.
- Query-Word Position Distribution — Where query words appear in the document. Front, body, title, headings, anchors. Position carries weight.
- Query-Word Density Pattern — How concentrated query words are in different parts of the document. Even distribution suggests broad coverage; concentrated distribution suggests targeted matching.
- Combined Rank Value — Weighted combination of relevance score and similarity score. The weights can be tuned per query class or learned from training data.
Key Insight: The patent foreshadows learning-to-rank by introducing the feature-vector representation of documents and queries as inputs to a learned combination function. Modern learning-to-rank pipelines use far richer feature vectors but the structural idea (rank from multi-feature similarity rather than single-score relevance) is exactly this patent's contribution.
<\/section>The Process
End-To-End Adaptive Ranking
The ranking pipeline runs on every query, producing per-document rank values from the relevance and similarity components.
- Query Arrival — User submits a query. The query is parsed and its feature vector is constructed.
- Candidate Retrieval — Standard inverted-index retrieval produces a candidate set of documents matching the query terms.
- Compute Relevance Scores — For each candidate, compute the relevance score using the classical relevance function.
- Compute Similarity Scores — For each candidate, construct its feature vector and compute similarity to the query feature vector.
- Combine And Rank — Combine the two scores per candidate into a rank value. Sort by rank value.
- Return Top-K — Return the top-ranked documents as the search results.
What This Means for SEO
What This Means for SEO
Adaptive ranking from feature vectors is the conceptual basis for modern learning-to-rank pipelines. The implications for content optimization span structure, query-word placement, and the kinds of features that the engine reads beyond pure term match.
- Structure Matters As Much As Content — Document attributes (headings, structure, length, freshness) feed directly into the feature vector. A well-structured page beats an equally relevant but unstructured one because the similarity score rewards structural alignment with query expectations.
- Query Word Position Is A Ranking Signal — Where query words appear (front, body, title, headings) is part of the feature vector. Front-loading important terms and placing them in headings beats burying them in body text.
- Density Distribution Matters — Query words evenly distributed across a long document signal broad coverage. Concentrated query words in one section signal focused matching. Match the distribution to the query type you target.
- Different Query Types Reward Different Features — Adaptive ranking can weight features differently per query class. Informational queries may weight structural completeness; transactional queries may weight clear calls-to-action. Content should match the feature weighting of its target query type.
- Freshness Is A Feature — Document freshness can enter the feature vector. For news, trends, and time-sensitive queries this signal weights heavily. Stale content competes worse on these queries even when relevance is strong.
- Long-Form Wins When The Query Demands Coverage — Document length is a feature. Long, comprehensive documents have feature-vector signatures that align with broad-coverage queries. Short documents win on focused queries. Match length to query type.