LLM Data Processing With Symbolic Verification for Output Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer applications struggle with processing heterogeneous and broad data collections, leading to impractical schema building and limited understanding of natural language, with deep learning models producing unreliable and inaccurate outputs.
Innovation Solution
Utilizing a structured, machine-readable representation of data, such as a universal language, to interact with large language models (LLMs) for improved output generation and validation, including methods to avoid hallucination and add citations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep learning models are used to process natural language, then the system can handle broad and heterogeneous data collections, but the output becomes unreliable and inaccurate
Solution Approach 1:
The patent introduces an intermediary verification system that acts as a mediator between the deep learning model and the final output. This intermediary layer validates predictions against structured knowledge bases and factual databases, correcting inaccuracies while preserving the model's ability to handle diverse data types.
Solution Approach 2:
The system implements feedback loops where model outputs are continuously validated against ground truth data and structured knowledge sources. Incorrect predictions trigger corrective feedback that adjusts the verification process, improving reliability without sacrificing the model's adaptability to different data formats.
2Adaptability or versatility
If a comprehensive schema is built to cover all possible data types, then the application can handle any heterogeneous data, but the schema becomes enormously complex and impractical to build
Solution Approach 1:
The patent employs a universal data representation framework that uses a small set of core schema elements capable of representing multiple data types through polymorphic structures. This universal schema can accommodate heterogeneous data including text, numbers, dates, and structured objects without requiring separate complex definitions for each data type.
Solution Approach 2:
The system dynamically adjusts schema parameters based on the specific data being processed. Rather than maintaining a static comprehensive schema, the framework selectively activates relevant schema elements and parameters according to the data type and context, reducing complexity while maintaining full data coverage capability.
3Productivity
If statistical machine learning is used to process natural language, then the system can learn from large datasets, but the results cannot be explained and the system lacks real understanding
Solution Approach 1:
The patent segments the processing pipeline into distinct stages: statistical learning for pattern recognition, structured knowledge representation for semantic meaning, and verification for factual accuracy. Each segment handles specific aspects of understanding, preserving semantic information that would otherwise be lost in purely statistical approaches.
Solution Approach 2:
Structured knowledge bases and semantic representation systems serve as intermediaries between statistical learning outputs and final interpretations. These intermediaries maintain explicit semantic relationships and factual constraints, preventing the loss of meaningful understanding while allowing statistical methods to drive learning from large datasets.
Data Source
AI summary
There is provided a method of improving the operation of a generative AI large language model (LLM)-based data processing system, by operating the LLM-based system in conjunction with a non-LLM data processing system; and in which (a) the LLM-based system sends a continuation as an input to the non-LLM system, and (b) the non-LLM system (i) uses symbolic representations to perform non-statistical reasoning on the input from the LLM-based system and (ii) generates a reasoned prompt or other context.