Privacy-Preserving Text Insight Mining via Dual Generative Model Bias Comparison

The patent segments the analysis into two distinct model types: a first generative language model trained on general internet text that provides baseline responses, and a second generative language model trained on specific domain text that provides domain-specific responses. By comparing outputs from these segmented models, the system recovers insight information that would otherwise be lost in bias-minimized models, while maintaining fairness through controlled comparison rather than raw biased outputs.

The patent introduces trained generative language models as intermediaries between raw user-generated content and analysis applications. The models are trained offline on large corpora including user-generated content, then deployed to generate responses to new inputs. This intermediary approach allows the system to leverage the adaptability and versatility of models trained on extensive data while protecting user privacy, as the trained models process information without requiring access to or storage of original user data.