Improving AI Language Understanding with Aligned Knowledge Graphs

This patent describes a method for making AI language models smarter by combining them with knowledge modules, which learn from two different, but related, knowledge graphs.

What it covers

This patent details a computer-implemented method to enhance how AI understands language by jointly training a language module with a knowledge module. First, it obtains two separate knowledge graphs, each containing entities (like people or places) and relations (how they connect), as described in Claim 1. It then aligns these two graphs, finding where entities and relations in one graph match those in the other. A language module, which creates numerical representations (embeddings) for words and concepts, and a knowledge module are accessed. The knowledge module is trained using these embeddings. Crucially, the language and knowledge modules are integrated so they can exchange information: the knowledge module gives 'knowledge information input' to the language module, and the language module provides 'context information input' to the knowledge module, as stated in Claim 1. This integrated system is then further trained using the aligned knowledge graphs to perform 'semantic analysis' for entities and relations, meaning it understands their meaning and relationships better. For example, Claim 4 describes using this system to translate speech from one language to another, leveraging the combined knowledge.

What it doesn't cover

• —Does not cover training a language model without explicitly integrating it with a separate, dedicated knowledge module.
• —Does not cover systems that use only a single knowledge graph, as it specifies obtaining and aligning a 'first knowledge graph' with a 'second knowledge graph'.
• —Does not cover language models where the knowledge module and language module do not actively exchange information in a two-way flow, providing input to each other.
• —Does not cover language understanding systems that do not perform 'semantic analysis for the entities and entity relations' based on learned knowledge.
• —Does not cover simple machine translation or speech transcription without the underlying joint training and knowledge graph alignment described.

The clever bit

The novelty lies in the 'joint training' and 'integration' of the language and knowledge modules, where they continuously feed information to each other. This creates a powerful feedback loop, allowing the language module to benefit from structured knowledge and the knowledge module to benefit from contextual language understanding, all while leveraging insights from 'multi-aligned' knowledge graphs.

Why it matters

Understanding language is a core challenge for AI. This patent offers a way to make AI systems much better at grasping meaning by giving them structured knowledge, not just statistical patterns from text. By combining language models with knowledge graphs, AI can move beyond just predicting the next word to truly understanding facts and relationships. This is vital for creating more accurate virtual assistants, search engines, and specialized AI tools that can reason about information.

Real-world examples

• 1.Advanced search engines that understand complex queries
• 2.Smart virtual assistants like Microsoft Copilot or Google Assistant
• 3.Enterprise AI systems for analyzing company-specific documents
• 4.Cross-lingual information retrieval systems
• 5.Automated content summarization tools