Vertical Selection
Definition
Predicting which specialised search services — verticals such as news, images, video, local, shopping or travel — are relevant to a query typed into a general search box, including the answer “none”, so their results can be blended into the main results page. The blended page is aggregated search; vertical selection is its first and decisive step.
It is the web-search ancestor of Query Routing, and a cooperative case of Federated Search resource selection: the search engine owns every vertical, so it has access to evidence an uncooperative federation never sees — above all, each vertical’s own query log.
Evidence
From Sources of Evidence for Vertical Selection (Jaime Arguello, Fernando Diaz, Jamie Callan, Crespo — SIGIR 2009):
| Evidence | Examples |
|---|---|
| Query string | Trigger words (“news”, “pictures”), geographic terms |
| Vertical query logs | Likelihood of the query under a language model of queries users typed directly into that vertical |
| Vertical corpora | Resource-selection scores (ReDDE, Clarity) over sampled vertical content, or Wikipedia proxies for text-poor verticals |
Query-log evidence was the strongest single predictor; a supervised classifier combining all three was best.
Properties Worth Keeping in Mind
- “No vertical” is the biggest class. About a quarter of queries — mostly navigational — should trigger nothing, so an explicit abstain class is part of the design, not a fallback.
- Error costs are asymmetric. An irrelevant vertical block may annoy users more than a missing one.
- Verticals are unlike each other. Some are genres (travel), some media types (images, video), some highly dynamic (news, where same-day demand matters).
Related Concepts
- Query Routing — the general pattern; vertical selection is its web-search instance
- Federated Search — resource selection, of which this is the cooperative case
- Query Classification · Search Intent
- Search Scopes — the user-chosen counterpart to engine-chosen verticals
Articles
- Sources of Evidence for Vertical Selection — the SIGIR 2009 paper defining the evidence sources
- Efficient Federated Search for RAG using Lightweight Routing — the same selection problem for RAG data sources