Selective Content Generation Using Unobjectionable Element Databases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing foundational models trained on large datasets often generate content that is objectionable due to inclusion of copyrighted, obscene, or prejudicial material, making it challenging to produce content that is safe for use.
Innovation Solution
A content generation system that uses a foundational model trained on both objectionable and unobjectionable data, where prompts are designed to select elements from an unobjectionable dataset to generate content, ensuring that the output does not include objectionable material by requesting specific elements that satisfy criteria related to a topic of interest, and verifying the correctness of the generated content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a foundational model is trained on a large dataset scraped from the internet, then the model can generate diverse content, but the generated content may include objectionable material such as copyrighted, obscene, or prejudicial content
Solution Approach 1:
The system segments the content generation process into multiple stages: (1) generating candidate elements from the foundational model, (2) filtering candidates through an objectionable content detector, and (3) selecting final content from approved candidates. This segmentation allows the system to maintain content diversity while systematically eliminating objectionable material through the filtering stage.
Solution Approach 2:
The patent introduces an intermediary component - the objectionable content detector - that sits between the foundational model and the final output. This detector acts as a mediator that evaluates generated content and blocks objectionable material before it reaches the user, allowing the foundational model to generate diverse content without directly exposing harmful outputs.
2Object-affected harmful factors
If prompts are designed to select elements from an unobjectionable dataset, then the generated content is free from objectionable material, but the system complexity increases due to additional verification steps
Solution Approach 1:
The system performs preliminary action by pre-curating an unobjectionable dataset and pre-training the foundational model to preferentially select from this safe dataset. By preparing the safe content pool in advance and building safety into the model's training, the system reduces the need for complex real-time filtering mechanisms, thereby lowering overall system complexity while maintaining content safety.
Solution Approach 2:
The patent implements feedback mechanisms where the objectionable content detector provides feedback to the prompt generation process. When objectionable content is detected, the system adjusts subsequent prompts to avoid generating similar content, creating a closed-loop system that learns from detection results and reduces complexity by adapting to patterns of safe content generation.
3Reliability
If the system verifies the correctness of generated content by checking against the unobjectionable dataset, then content quality is ensured, but the processing time increases
Solution Approach 1:
The system applies partial verification by checking only the most critical aspects of generated content against the unobjectionable dataset rather than performing exhaustive verification of all content properties. This selective verification approach ensures content quality for the most important safety criteria while reducing processing time by avoiding unnecessary checks on already-satisfied conditions.
Solution Approach 2:
By pre-organizing the unobjectionable dataset with efficient indexing and metadata structures during the preparation phase, the system enables rapid verification during content generation. The preliminary organization of safe content with searchable characteristics allows the verification process to quickly determine whether generated content meets quality standards without time-consuming full comparisons.
Data Source
AI summary
In one implementation, a method of generating content is performed at a device having one or more processors and non-transitory memory. The method includes obtaining a database including a plurality of elements. The method includes generating a prompt to select one of the plurality of elements that satisfies one or more criteria regarding a topic of interest. The method includes submitting the prompt to a foundational model trained on a training dataset. The method includes receiving, from the foundational model in response to the prompt, content including a selected one of the plurality of elements.