We recently released a demo of our car model intent model (too many machine learning models mixing with other models!). Often when people search they don’t use the ‘correct’ words, they misspell them or they leave out important query terms, because they know what they mean. We translate keywords into an intent, which makes all the implicit assumptions explicit. The demo is for just one part of the intent model that focuses on matching car makes and models. We use this more generally to match all kinds of entities in our knowledge graphs for fashion, homeware, electronics, automotive
We use intents internally when we match keywords to each other (in related searches), keywords to pages or vice versa (when we create new pages or hide pages without demand), or pages with each other (in deduplication). For each of these, we express the underlying meaning as explicitly as possible as an intent.
If you’re not sure whether the model got the right result, you can ask the SERPs and Google it!
The car intent model handles
- simple queries, like audi a3 2.0 tfsi dsg 2005-> audi/a3 orbmw 118i m sport-> bmw/1-series.
- misspelt queries, like range river defendr-> land rover/defender
- mixed queries, like mercedes jeep-> mercedes-benz/g-series,vw van-> volkswagen/transporter,ford estate-> ford/focus, oraudi suv-> audi/q7
- queries without car models: sometimes it will assume a default model, like abarth -> abarth/500, or only return the make
- a little imagination! Sometimes it does weird things, like ford cougar-> ford/kuga. Ford did have a Cougar but the Kuga is a lot more popular, and the model assumes the user meant that. (Asking the SERPs gives the same answer).