LLM Document Compliance Evaluation With Hallucination Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The creation and evaluation of textual documents, such as statements of work, can be tedious and prone to errors due to the need for knowledge and experience to ensure compliance with specified rules and guidelines, and the potential for hallucinations generated by large language models.
Innovation Solution
A method involving a large language model to generate and evaluate document chunks for compliance, using a similarity score and hallucination score to ensure coherence and accuracy, with separate models for standard and non-standard chunks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a large language model is used to generate document chunks, then document generation speed is improved, but hallucinations and compliance errors may occur
Solution Approach 1:
The system implements a feedback mechanism where the second large language model evaluates the output of the first model, checking for hallucinations and compliance issues. The evaluation results are used to identify and correct errors in the generated document chunks, ensuring that the final output meets compliance requirements while maintaining generation speed.
Solution Approach 2:
A second large language model is introduced as an intermediary evaluator between the document generation process and the final output. This intermediary model specifically checks for hallucinations and compliance issues without being involved in the generation process itself, allowing the system to maintain high generation speed while ensuring reliability through specialized evaluation.
2Reliability
If manual evaluation of documents is performed, then compliance accuracy is improved, but time consumption increases
Solution Approach 1:
The patent replaces manual mechanical evaluation with an automated large language model-based evaluation system. The second model performs compliance checking and hallucination detection automatically, achieving accuracy comparable to manual review while dramatically reducing the time required for evaluation.
3Reliability
If document compliance checking is performed thoroughly, then compliance accuracy is improved, but processing complexity increases
Solution Approach 1:
The compliance evaluation process is segmented into distinct functional components: the first large language model handles document generation, the second model handles evaluation and hallucination detection, and templates provide compliance criteria. This segmentation allows each component to specialize in specific tasks, improving overall compliance accuracy while managing system complexity through modular design.
Data Source
AI summary
A method of evaluating a first document for compliance is disclosed herein that includes separating the first document into multiple chunks of text and evaluating a first chunk to determine whether the first chunk is a standard chunk type or a nonstandard chunk type. In response to the first chunk being a standard chunk type, the method can include determining a purpose of the first chunk, retrieving a first template chunk having a similar purpose, and determining a first similarity score representative of a similarity between the first chunk and the first template chunk. In response to the first chunk being a nonstandard chunk type, the method can include providing the first chunk to a first large language model, determining (by the first large language model) a first hallucination score, and formulating a compliance score dependent on the first similarity score or the first hallucination score.