Domain-Specific GAN Discriminator for AI Hallucination Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generative AI solutions, such as chatbots, often 'hallucinate' by providing incorrect information, especially when answering specific questions about a company's products or policies, leading to inefficiencies and inaccuracies.
Innovation Solution
A generative adversarial network (GAN) is employed, comprising a generative network and a discriminative network, where the discriminative network is trained on domain-specific information (e.g., retirement, cyber, legal, compliance, human resources, privacy, fairness) to reduce hallucinations by evaluating the generated data and determining its accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a general generative AI model is used to answer customer questions, then the system can handle general inquiries, but it hallucinates when asked specific questions about company products or policies
Solution Approach 1:
The patent divides the single generative AI model into two separate specialized models: a general-domain generative model for handling general inquiries and a domain-specific generative model trained on company product and policy data for handling specific questions. This segmentation allows each model to excel at its designated task, preventing hallucinations on specific domain questions while maintaining versatility for general queries.
Solution Approach 2:
The patent introduces a classifier as an intermediary component that receives user questions and determines whether to route them to the general-domain model or the domain-specific model. This mediator enables the system to adaptively select the appropriate model based on question characteristics, ensuring reliable answers for specific domain questions while maintaining the ability to handle general inquiries.
2Adaptability or versatility
If the generative AI model is trained on general data, then it can respond to various topics, but it provides incorrect information about specific company products
Solution Approach 1:
The training data and model capabilities are segmented into two distinct domains: general knowledge covered by the general-domain model trained on broad data, and specific company product information covered by the domain-specific model trained on proprietary data. This segmentation ensures measurement precision for product information while preserving adaptability across topics through the general model.
Solution Approach 2:
The patent applies local quality by creating a domain-specific model with specialized training on company product and policy data, giving it superior local knowledge quality for specific topics. Meanwhile, the general-domain model maintains broad topical coverage. The classifier routes queries to the appropriate model based on the required quality level for that specific topic area.
Data Source
AI summary
The following relates generally to generative artificial intelligence (AI), and more particularly to reducing “hallucinations” in generative AI solutions. In some embodiments, one or more processors: (i) receive an input statement; and (ii) generate a response to the input statement by inputting the input statement into a general adversarial network (GAN), the GAN comprising: (a) a generative network configured to send and receive data to a discriminative network; and (b) the discriminative network configured to send and receive data to the generative network, wherein the discriminative network was trained based on information of at least one domain.