Google patent: subjective entity attributes

In May 2022, Google was granted the patent "Identifying subjective attributes by analysis of curation signals" (US 11,328,218). It describes how search systems derive subjective traits of entities—such as "funny," "cute," or "awesome"—from reactions on the web and predict them for new content. For SEO teams, this is more than patent trivia: it links entity understanding, user-generated content (UGC), and machine learning in a clear signal model.

Why subjective attributes matter to Google

Objective facts about an entity—name, date, address—are not enough for modern search. Users often describe media, brands, or places emotionally and subjectively. The patent closes that gap: it defines a vocabulary of subjective attributes and maps them to entities ranging from media clips and blog posts to people and businesses. Relevance scores between 0.0 and 1.0 measure how strongly an attribute fits an entity—for example when "cute" appears in half of all comments on a video.

UGC as the core data source

A critical building block is user-generated content. Google cites comments on social networks, reactions to blog posts, product and movie reviews, and non-textual signals such as like, dislike, +1, sharing, bookmarking, or playlist entries. These curation signals supply the raw data from which the system extracts initial attribute sets for an entity. As UGC spreads, this patent logic becomes more important for evaluating content without classic keyword density.

• Textual comments and reviews
• Likes, shares, and bookmarks
• Playlist and feed visibility
• Weighting by user authority

From signal to trained classifier

The process follows a repeatable pattern: first, a "Subjective Attribute Identifier" extracts attributes from user reactions. A "Relevancy Scorer" calculates frequencies and drops weak hits—often only the k strongest attributes remain. In parallel, a feature extractor builds vectors from color, texture, audio, word frequencies, or metadata. Google trains classifiers such as support vector machines, AdaBoost, neural networks, or decision trees on these input-output pairs. After training, attributes can be predicted for new entities without comments—such as freshly uploaded videos or articles with no engagement yet.

Patent facts at a glance

System architecture and data stores

The described architecture includes servers, an entity store, and client devices over public or private networks. The entity store holds media, web pages, and reviews; a subjective attribute manager ties together web server, identifier, scorer, feature extractor, and classifier. Users can choose which data is collected, per the patent—a nod to privacy and transparency when using attribute-based personalization.

Inverse mappings and search use cases

After processing many entities, an inverse mapping emerges: which objects carry the attribute "funny"? That speeds retrieval for keyword search, playlist filling, ad delivery, and new training data. For publishers, tags and descriptions that mirror user language can support findability in attribute-based surfaces—not only through classic meta keywords.

Physical entities via cyber proxies

Restaurants, actors, or local businesses often lack direct UGC fields on their own domains. The patent allows evaluation via fan pages, review portals, or aggregated feed signals; subjective attributes still attach to the real-world entity. For local SEO and reputation management, that is central: third-party platform signals feed Google's entity model, not only on-page copy.

Relevancy scores and attribute vocabulary

The patent details how relevancy scores are built: frequency in comments, weighting of individual users, and thresholds decide which attributes stick to an entity. A score between 0.0 and 1.0 can reflect the share of comments containing a term like "awesome." Weak attributes are dropped or zeroed; often only the k strongest traits remain. The vocabulary itself can start manually and grow via NLP from new reactions—including hierarchical meta-attributes such as "positive" or "negative" for sentiment words.

Feature vectors beyond text

The classifier is not fed by language alone. For video and images, color, texture, and intensity matter; for audio, amplitude and spectral coefficients; for text documents, word frequencies, sentence length, and formatting. Metadata and external processing libraries can precompute features. That allows subjective labels to be estimated even with little UGC yet—a relevant case for newly published guides, product pages, or local listings without review history.

Human validation and retraining

A second method suggests predicted attributes to the uploader; users confirm, add, or remove suggestions. Removed attributes are stored as negative examples and flow into retraining—for example every 100 loop iterations or on a fixed schedule. The model stays tied to real curation instead of automatic text matches alone.

Practice for SEO and content strategy

Anyone optimizing for visibility should treat UGC not only as social proof but as training input for subjective entity labels. Authentic comments, consistent review language, and clear reaction signals can strengthen attribute assignment. Structured feature data—metadata, media quality, text statistics—also pays off because the classifier works without UGC. The patent underscores that Google's entity SEO moves beyond Knowledge Graph facts toward crowd-based semantics and predictable emotional dimensions.

Inventors Hrishikesh Aradhye and Sanketh Shetty, with assignee Google LLC, anchor the invention firmly in the search ecosystem. Teams that use patents as early indicators should align comment quality, engagement signals, and semantic tags more closely with entity goals—whether the focus is news, video, or local brands.