What is attribute extraction?
Attribute extraction pulls structured values, such as color, material or size, out of unstructured text or images and stores them as fields on a product record.
In plain words: A seller writes 'olive cotton jacket with buttons'; attribute extraction turns that into Color: Olive, Material: Cotton and Closure: Buttons. The input can be a title, a description, a specification sheet or a product photo. Careful extraction leaves a field empty when the evidence does not support a value, instead of guessing one.
Also called: attribute tagging, product attribute extraction
Published by Youzu · Editorial policy
- Published

- Collar: point collar. Visible in the photo, so it can be read.
- Color: rust red. Read from the photo and kept to your color list.
- Pockets: 2 flap chest pockets. Visible in the photo.
- Material: left empty. A photo cannot show what the fabric is made of; the value waits for the label or the seller.
- Collar: point collar. Visible in the photo, so it can be read.
- Color: rust red. Read from the photo and kept to your color list.
- Pockets: 2 flap chest pockets. Visible in the photo.
- Material: left empty. A photo cannot show what the fabric is made of; the value waits for the label or the seller.
Why it matters
Filters, comparison tables and channel feeds read fields, not sentences. When facts stay inside free text, a shopper filtering by material never sees the product, and a feed that requires a color rejects it. Extraction is usually the first step of enrichment, because every later check needs typed values to compare.
How it works
Start from the category
The category decides which attributes exist, so extraction starts from a category node or a category suggestion.
Read the evidence
Text and images are read for values of those attributes, and each value notes where it came from.
Check and store
Values are checked against allowed values and types, then stored as fields; unsupported fields stay empty for review.
| Field | Value | Evidence |
|---|---|---|
| Product type | Jacket | Suggested from your category |
| Color | Olive | Suggested from your facts |
| Material | Cotton | Suggested from your facts |
| Closure | Buttons | Suggested from the photo |
| Fit | Left empty | Not visible in the photo. |
| Size | Left empty | Ask the seller. |
Examples
Text and photo together
From the typed facts 'jkt', 'olive' and 'cotton' and a photo, the jacket gets Color: Olive, Material: Cotton and Closure: Buttons. Fit and size stay empty, because neither the facts nor the photo show them.
A number with its unit
The supplied fact '2000 mm' becomes a width field with its unit, so it can be filtered and compared instead of being read inside a sentence.
Facts that stay with one item
In a classifieds listing for a used walnut chair, condition, location and the note about a small scuff come from the seller and stay with that item. The width is left empty with a prompt to ask the owner.
Common mistakes
Guessing to fill a field. A guessed material passes validation, then fails the shopper who filtered on it. Leave unsupported fields empty.
Extracting before the category. Values read before the category is known land in the wrong fields, or in fields the category does not have.
Related terms
- Product data enrichment
- Adding missing information or improving existing information on a product record so products can be found, filtered and compared.
- Advisory suggestion
- An advisory suggestion is a proposed value that the receiving application may apply, edit or ignore; it changes nothing by itself.
- Listing quality
- How complete, consistent and policy-compliant a listing is, judged from its text, attributes, images and category.
In Youzu
Youzu helps marketplaces, classifieds and retailers turn raw seller and supplier product data into structured, reviewable catalog records.
Sources
Product data specification, Google Merchant Center Help. support.google.com/merchants/answer/7052112. Accessed .
Product, Schema.org. schema.org/Product. Accessed .
FAQ
Questions buyers ask
Not quite. Attribute extraction pulls values out of text or images you already have. Enrichment is the wider job of adding or improving information on a record, which can include extraction, category mapping, better titles and descriptions, identifiers and images. Extraction is usually one step inside an enrichment workflow.
Partly. A photo can show color, pattern, shape or closure, but not material composition, care instructions or measurements. Treat photo-derived values as suggestions to review, combine them with the seller's text where you can, and leave a field empty when neither source supports a value.
The ones your categories define. Youzu maps products to your category tree and extracts the attributes your category defines; numbers and yes/no values stay typed, and values the evidence does not support are left empty. Send a sample file and your category tree to see it on your own products.
Last reviewed
Related
Keep reading
Product
Extract product attributes from images
Read color, material and style from product photos into typed attributes for filters and search; empty when the image does not support it.
Glossary
What is product data enrichment?
Product data enrichment defined in plain words, with a worked example, common mistakes and links to the guides and Youzu products that apply it.
Product
AI product data enrichment from the text and images you already have
Youzu fills the attributes your categories define from text and images, localizes titles and descriptions and leaves a field empty when evidence is thin.