How Computers Build Knowledge Maps from Text Documents

Adobe's patent describes a system that automatically builds a structured map of information, called a knowledge graph, by identifying key concepts (noun-phrases) and their connections within written text, then using this map to modify online platforms.

What it covers

This patent describes a method for creating a knowledge graph from a collection of text documents, referred to as a 'text-based knowledge corpus'. First, the system extracts individual sentences from the text. For each sentence, it identifies 'noun-phrases' and determines the relationships between them using both the sentence's grammatical structure ('dependency paths') and a 'relation classifier' (Claim 1). This classifier can predict connections between any pair of noun-phrases, not just pre-defined 'named entities' (Abstract). Based on these noun-phrases and their relations, a small 'sub-graph' is created for each sentence. Finally, all these sub-graphs are merged together to form a complete knowledge graph for the entire text collection (Claim 1). For example, if a sentence says 'Adobe developed Photoshop', the system would identify 'Adobe' and 'Photoshop' as noun-phrases, determine the 'developed' relation between them, and add this to the graph. The resulting knowledge graph can then be used to identify specific entities or relations for modifying an online platform (Claim 1).

What it doesn't cover

• —Does not cover methods that use other types of phrases (like verb phrases or adjective phrases) to build the graph, as it specifically claims using 'noun-phrases in the sentence without using other types of phrases in the sentence' (Claim 1).
• —Does not cover determining relations solely based on statistical co-occurrence without considering 'dependency paths of the sentence' or relations 'predicted by a relation classifier' (Claim 1).
• —Does not cover knowledge graphs where entities are restricted to a predefined set of named entities, as the relation classifier is 'configured to predict a relation between a pair of entities without restricting the entities to a set of named entities' (Abstract).
• —Does not cover systems that don't merge sentence-level sub-graphs to form the final knowledge graph (Claim 1).
• —Does not cover systems that don't transmit identified entities or relations to an 'online platform' for 'modifying the online platform' (Claim 1).

The clever bit

The cleverness lies in combining grammatical analysis ('dependency paths') with a machine learning model ('relation classifier') to find connections between *any* noun-phrases, not just pre-identified 'named entities'. This allows for a more flexible and comprehensive understanding of text. Building the larger graph by merging smaller, sentence-level 'sub-graphs' is also a key innovation, enabling a scalable approach to knowledge extraction.

Why it matters

Knowledge graphs are crucial for helping computers understand and organize vast amounts of information, much like how humans connect ideas. This patent offers a way to automatically build these structured maps from unstructured text, which is a fundamental challenge in artificial intelligence. By extracting precise relationships between concepts, it can power more intelligent search, content recommendations, and automated content generation across various online services.

Real-world examples

• 1.Google's Knowledge Graph
• 2.Wikipedia's structured data (Wikidata)
• 3.Content management systems
• 4.Semantic search engines
• 5.Recommendation systems for online platforms
• 6.AI-powered assistants that answer factual questions