Recognizes when a non-question-shaped query is actually expressing a question, so it can be routed to the answer-passage pipeline that powers featured snippets.
Patent Overview
- Inventor
- Steven D. Baker
- Assignee
- Google LLC
- Filed
- 2016-04-26
- Granted
- 2018-02-20
- Application Number
- US 15/138,907
The Challenge
Question Intent Without Question Words
Many information-seeking queries are not phrased as questions. "Capital France" carries the same intent as "What is the capital of France?" but lacks question words. Routing the second to answer-passage selection and the first to plain retrieval misses opportunities to serve direct answers to users. A robust question classifier needs to recognize implicit question intent without relying on question-word presence.
- Question Words Are Not The Only Signal — Filtering for "what", "why", "how", "who", "when", "where" misses a large fraction of question-intent queries that omit those words. Mobile and voice queries especially trend toward keyword form.
- Surface Form Hides Intent — Bare noun-phrase queries like "Eiffel Tower height" are unambiguously seeking a specific fact, even though they look like keyword soup. The intent is in the entity-attribute pattern, not in the words used.
- Need A Template Match That Recognizes Implicit Questions — The system needs to recognize structural patterns common to implicit questions, distinct from generic exploratory queries. Template matching captures structural regularity without requiring lexical cues.
- Single-Template Matching Misses Compound Patterns — Some implicit questions match multiple shallow templates that individually do not indicate question intent. Combining template matches (single-template plus second-template) is needed to catch them.
- False Positives Waste Answer Slot — Treating exploratory queries as questions and routing them to answer-passage selection wastes the answer slot on bad fits. The classifier must be precise enough that promotion to the answer pipeline is consistently rewarded.
Innovation
Two-Stage Template Matching
The query is compared against a library of templates. The first match identifies the surface structure of the query (e.g., entity-attribute). A second template is then derived from the first; if the second template is indicative of a question query, the original query is classified as an implicit question.
- Receive Query In Unstructured Form — A free-form query arrives without explicit question markers. The classifier does not assume any particular shape.
- Match Against Query Templates — Compare the query terms to a set of known query templates. Find the best matching template for the query's structure. Templates capture syntactic shapes and entity-type patterns.
- Check The First Template — If the first matching template is already indicative of a question query, return question classification directly. Some templates are themselves strong question signals.
- Derive A Second Template — If the first template is not question-indicative on its own, derive a second template from it. The derivation captures more specific structural variants or related patterns.
- Classify Based On Second Template — If the second template is indicative of a question query, classify the original query as an implicit question. The chained match is what catches the compound patterns single-template matching misses.
- Route Accordingly — Implicit question queries are routed to the answer-passage pipeline so they can compete for featured snippets and direct answers. Non-question queries continue through standard retrieval.
Templates As Intent Signatures
The patent's contribution is treating query templates as multi-stage intent signatures rather than single-stage classifications. Some templates are not question-indicative on their own but become question-indicative when chained to a derived template.
Chained Template Matching
If a query matches template T1 and T1 derives to T2 which is question-indicative, the query is treated as an implicit question. The chain catches what single-stage matching would miss.
- Direct Question-Indicative Templates — Some templates (e.g., "what is X", "how to Y") are themselves question signals. Queries matching them are classified directly.
- Derived Question Templates — Other templates (entity-attribute, comparison) are not question signals on their own. The system derives related templates that are question signals, and chains the classification through.
Intent recognition is not a single match. It is a chain of structural inferences.
<\/section>Technical Foundation
Query Templates As Intent Signatures
Templates describe structural patterns of queries. Each template has a position-aware structure and an associated intent classification.
- Query Template — A pattern that captures the syntactic shape of a query, including position-specific entity types and optional question markers. Templates can be hand-curated or learned from labeled data.
- Question-Indicative Template — A template that is empirically associated with question-intent queries based on offline analysis. Some templates carry this label; others do not.
- Template Derivation — A mapping from one template to a related template, used to chain shallow matches into deeper intent classification. The derivation can be hand-defined or learned.
- Chained Classification Outcome — The final classification combines the first-template match with the derivation result. A query is an implicit question if either the first template or the derived second template is question-indicative.
Key Insight: The chaining is what makes the classifier expressive without exploding the template library. A small set of templates plus a derivation graph can cover a large variety of implicit question shapes, where a flat single-stage classifier would need a much larger template inventory to achieve similar coverage.
<\/section>The Process
Classification Pipeline
The classifier runs early in query processing, before retrieval. The classification result determines whether the query is routed to the answer-passage pipeline.
- Receive Query — A free-form query arrives without classification metadata.
- First-Template Match — Compare the query against the template library. Identify the best-fitting template.
- Direct Classification Check — If the matched template is itself question-indicative, classify the query as an implicit question and route to the answer pipeline. Done.
- Derive Second Template — If the first template was not question-indicative, use the derivation graph to find a related second template.
- Second-Template Classification Check — If the second template is question-indicative, classify the query as an implicit question via the chain. Otherwise, treat as standard retrieval.
- Route Query — Implicit questions go to the answer-passage scoring pipeline. Non-questions proceed through standard retrieval without answer-slot consideration.
What This Means for SEO
What This Means for SEO
Implicit question detection is one of the upstream steps that determines whether your page gets a chance to be a featured snippet or direct answer. Knowing the rule changes how you should structure content for question intent and how broadly you should target question-shaped queries.
- Bare Noun-Phrase Queries Compete For Answers Too — Targeting "Eiffel Tower height" is not less competitive than targeting "how tall is the Eiffel Tower". Both queries can route to the answer-passage pipeline. Optimize for both forms and treat them as one intent.
- Structure Content For Question Intent Even Without Question Words — Direct-answer formatting (one-sentence answers near the top, fact tables, schema markup) helps even when the target query is a bare phrase. The classification happens upstream; your job is to be ready when it does.
- Map Your Topic's Implicit-Question Patterns — For each topic, list the entity-attribute shapes users might search ("X height", "X capital", "X cost", "X release date"). Each is an implicit question candidate worth covering.
- Entity-Attribute Pages Are Answer Opportunities — Pages built around an entity (a person, product, place, concept) with explicit attribute coverage (height, weight, cost, founded, capital) are aligned with the implicit-question template patterns. They consistently get answer-slot consideration.
- FAQ Sections Help Even For Non-FAQ Queries — FAQ-formatted sections cover both explicit and implicit question patterns. The Q-and-A shape is template-friendly even when the user query is a bare phrase.
- Voice Queries Are Often Explicit Questions — Voice search inputs tend to use explicit question words because people speak in full sentences. Cover both explicit and implicit forms for the same intent so both query shapes route to your page.
- Comparison Queries Are Implicit Questions Too — "X vs Y" patterns are implicit questions that the chained template logic can catch. Comparison content should be structured as if answering an explicit comparative question.