How a Computer System Improves Large Language Model Output with Structured Data

This patent describes a method where a separate computer system analyzes and refines the initial output from a large language model (LLM) by translating it into a highly structured, machine-readable language to reduce bias and improve accuracy.

What it covers

This patent details a computer-implemented method for improving the output of a Large Language Model (LLM). First, the LLM processes input data and generates an initial output (claim 1a). This initial output is then sent to a separate 'processing system' (claim 1c). This processing system is unique because it uses a highly structured, machine-readable language. This language represents data using 'semantic nodes' (specific meanings) and 'semantic links' between them, where even the links are semantic nodes (claim 1b). It also represents 'reasoning steps' and 'computation units' within this language (claim 1b). The processing system analyzes the LLM's initial output by translating it into this structured language and using its built-in reasoning steps and computation units (claim 1d). For example, if an LLM generates a response that contains factual errors or shows bias, this processing system can identify those issues by analyzing the translated output for factual inaccuracies (claim 4) or bias (claim 9), and then generate a corrected, 'bias-reduced version' or 'improved version' to present to the user (claim 1d).

What it doesn't cover

• —Does not cover systems that refine LLM output solely by using another LLM as a critic or editor, without translating the output into the specified structured, machine-readable representation for analysis.
• —Does not cover methods that improve LLM output using simple keyword filters or statistical models to reduce bias, without the explicit use of semantic nodes, semantic links, reasoning steps, and computation units.
• —Does not cover systems that provide additional context or instructions to the LLM *before* it generates its initial output, as the core claim focuses on analyzing the LLM's output *after* it has been generated.
• —Does not cover systems that only check for grammatical errors or stylistic improvements in LLM output, without also analyzing for factual inaccuracies, contradictions, or logical consistency using semantic reasoning.
• —Does not cover systems that use a structured language if that language does not represent semantic nodes, semantic links, reasoning steps, and computation units as explicitly defined in the claims.

The clever bit

The clever part is using a separate, highly structured processing system that translates LLM output into a machine-readable language with explicit semantic nodes, links, reasoning steps, and computation units. This allows for a robust, systematic analysis of the LLM's output for issues like bias or factual errors, rather than just relying on another AI or simple rules.

Why it matters

As Large Language Models become more common, ensuring their outputs are accurate, consistent, and unbiased is crucial. This patent addresses a core challenge in AI: making LLMs more reliable for real-world applications. By providing a structured way to post-process and validate LLM responses, it helps build trust and expands the types of tasks where LLMs can be safely deployed. This approach aims to make LLM-generated content more dependable for users.

Real-world examples

• 1.AI assistants providing fact-checked information
• 2.Content generation platforms with built-in accuracy checks
• 3.Automated research tools summarizing documents without bias
• 4.Customer service chatbots offering logically consistent responses