Model Authenticity Evaluation Using Dynamic Entity-Based Questions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing model authenticity evaluation methods, particularly for large models, face challenges in constructing high-quality datasets, suffer from sparse coverage scenarios, and are difficult to dynamically update, leading to inefficient and ineffective hallucination detection.
Innovation Solution
A method involving the extraction of named entities from input question data, transformation into new question data, and comparing model responses to determine authenticity, without relying on static question-answer datasets or data-driven tools, enabling dynamic and real-time evaluation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If static question-answer datasets are used for model authenticity evaluation, then dataset construction is simplified, but coverage of scenarios becomes sparse and detection quality deteriorates
Solution Approach 1:
The patent transforms the static question-answer dataset into a dynamic evaluation system. Instead of relying on pre-defined static questions, the system dynamically generates evaluation questions by extracting named entities from input questions and constructing new questions on-the-fly. This dynamic approach enables comprehensive scenario coverage while maintaining ease of implementation through automated question generation.
Solution Approach 2:
The system enables the model to evaluate itself by using its own capabilities to generate evaluation questions and answers. The model extracts named entities, constructs new questions, generates answers, and then evaluates its own responses against the generated questions. This self-service mechanism eliminates the need for extensive manual dataset construction while achieving comprehensive scenario coverage.
2Extent of automation
If data-driven detection tools are employed, then detection automation is improved, but detection quality suffers due to low quality of question-answer data
Solution Approach 1:
The system employs the model's own capabilities to generate high-quality evaluation questions and answers, eliminating dependence on external low-quality datasets. By leveraging the model's named entity extraction, question generation, and answer generation capabilities, the system creates quality evaluation data dynamically, ensuring both automation and detection quality.
Solution Approach 2:
The system implements a feedback mechanism where the model's responses are evaluated against dynamically generated questions and answers. The evaluation results provide feedback on model authenticity, which can be used to improve the model's performance. This closed-loop feedback ensures high detection quality by continuously validating the model's responses against dynamically generated ground truth.
3Device complexity
If discrete static question-answer data is used, then implementation is simplified, but real-time dynamic detection capability is lost
Solution Approach 1:
The patent implements real-time dynamic detection by replacing static question-answer data with dynamic question generation. The system extracts named entities from input questions, constructs new evaluation questions dynamically, and generates answers in real-time. This dynamic approach enables the system to adapt to any input question without relying on pre-defined static datasets, achieving both real-time detection and comprehensive scenario coverage.
Solution Approach 2:
The system performs preliminary actions by pre-defining the question generation framework and named entity extraction rules. Instead of having static questions ready, the system prepares the computational framework that enables rapid dynamic question generation. This preliminary setup allows the system to quickly generate evaluation questions and answers in real-time without complex runtime processing.
Data Source
AI summary
Embodiments of this specification disclose model authenticity evaluation methods, apparatuses, and devices. The method includes: obtaining first question data used to perform authenticity evaluation on a target model, and inputting the first question data to a target model to obtain a first response result corresponding to the first question data; extracting a named entity included in the first question data, and constructing second question data based on the named entity and the first question data, where the second question data is used to trigger the target model to output an analysis basis and a result for the first question data; inputting the second question data to the target model to obtain a model prediction result corresponding to the second question data; and determining an authenticity evaluation result of the target model based on the first response result and the model prediction result.