Decomposes an answer-seeking query into elemental components (question type plus answer types), then constructs answer elements that align element-to-element with the question, supporting structured direct-answer generation.
Patent Overview
- Inventor
- Nitin Gupta
- Assignee
- Google LLC
- Filed
- 2018-05-09
- Granted
- 2020-03-17
- Application Number
- US 15/975,373
The Challenge
Answers Need Structure That Matches The Question
A question query has internal structure: it has a question type (definitional, factual, comparative, instructional) and one or more expected answer types associated with that question type. A direct answer should mirror this structure: the answer types should match the expected answer types. Without explicit decomposition, answer generation produces unstructured prose that may or may not align with what the question asks for.
- Question Type Drives Answer Shape — Different question types require different answer shapes. The system has to identify the question type first so it knows what answer to construct.
- Multiple Answer Elements Per Question — A single question can require multiple answer elements (entity, attribute, value, qualifier). Decomposing the answer into elements ensures coverage.
- Element-Wise Alignment — Each answer element should correspond to part of the question. Element-wise alignment makes the answer auditable and verifiable.
- Search Results Are The Source — Answer elements are constructed from search results that satisfy the query. The results provide the raw material; the decomposition organizes it into elements.
Innovation
Decompose Question Into Types, Construct Answer Elements
The method receives a query with multiple terms, classifies it as an answer-seeking query of a particular question type, and obtains the answer types associated with that question type. The system then obtains search results satisfying the query and constructs answer elements that align with the expected answer types, producing a structured answer rather than unstructured prose.
- Receive Query — Multi-term query arrives at the system.
- Classify As Answer-Seeking — Upstream classifier flags the query as answer-seeking. Pure exploratory queries skip this pipeline.
- Identify Question Type — Classify the question type (definitional, factual, comparative, instructional, etc.). The type determines which answer types are expected.
- Obtain Expected Answer Types — From the question type, retrieve the associated answer types. A factual 'when did' question expects a date; 'how tall' expects a numeric height; 'who is' expects a person name plus context.
- Run Search To Get Results — Standard retrieval produces results satisfying the query.
- Construct Answer Elements — Extract content from search results that matches each expected answer type. Each extracted element corresponds to a slot in the structured answer.
- Assemble Structured Answer — Combine the answer elements into a coherent structured answer. The structure matches the question's expected shape.
Element-Wise Question To Answer Alignment
The patent decomposes both question and answer into structural elements that map to each other. This produces verifiable, structured answers rather than the free-form prose that pure generation can produce.
Type-Driven Decomposition
Question type implies expected answer types. The structured answer fills each expected type with content from search results.
- Question Type Classifier — Maps the input query to one of the known question types. Drives the rest of the pipeline.
- Expected Answer Types — Per-question-type list of the elements the answer should contain.
- Element Extraction From Results — Search-result content is parsed to extract per-element fillers.
What This Means for SEO
What This Means for SEO
Element-wise answer construction shapes how factual content gets surfaced in direct-answer modules. Knowing the type-to-element mapping informs how to structure factual content for extraction.
- Match Answer Shape To Question Type — For 'how tall' content, include a numeric height with units in a clearly labeled position. For 'when did' content, include a date in canonical format. The system extracts elements; obvious formatting helps.
- Structured Data Aids Element Extraction — Schema markup, structured data, and table formatting make element extraction more reliable. Plain prose works but structured forms work better for type-driven answer construction.
- Comprehensive Factual Pages Win Multi-Element Answers — Some answers have multiple elements (entity + attribute + qualifier). Pages that cover all expected elements get pulled into structured answers more reliably than partial pages.