Language Model Answer Verification System

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Users may be misled by incorrect information output from language models, as the accuracy of the information is not always correlated with the writing style, leading to potential misinformation.

Innovation Solution

An information processing system that acquires a question from a user, retrieves answer basis information from a database, requests a language model to generate an answer, and determines whether the model answer matches the retrieved information. The system then generates an output answer that includes the model answer and additional information indicating its reliability based on the match determination.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If a language model is used to generate answers, then natural sentence output capability is improved, but information accuracy deteriorates

Engineering Contradiction:
Improvenatural sentence output capabilityVSAvoidinformation accuracy
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The patent introduces an intermediary verification mechanism that compares the language model's output against a knowledge base. The system retrieves relevant information from the knowledge base and determines whether the model answer matches the retrieved information, acting as a mediator to verify accuracy before presenting the answer to the user.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system implements feedback by providing reliability information to the user about the generated answer. The answer generation unit outputs not only the model answer but also information indicating the reliability of that answer based on the match determination, allowing users to assess the trustworthiness of the information.

Inventive Principle:
Principle #23Feedback

2Productivity

If language model output is provided without verification, then response speed is improved, but misinformation risk increases

Engineering Contradiction:
Improveresponse speedVSAvoidmisinformation risk
Core Design Contradiction:
ProductivityVSObject-affected harmful factors

Solution Approach 1:

The system performs preliminary verification by retrieving information from the knowledge base and comparing it with the model answer before presenting the final output. This preliminary action ensures that the answer is verified against reliable sources while maintaining efficient processing through parallel operations.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The verification mechanism serves as an intermediary that quickly assesses the reliability of the model answer by comparing it with retrieved knowledge base information, preventing misinformation from reaching the user without significantly delaying the response.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentUS20250139376A1Information processing system, information processing method, and non-transitory computer readable storage medium
Publication Date: 2025.05.01 RAKUTEN GROUP INC
  • US20250139376A1 patent drawing
  • US20250139376A1 patent drawing
  • US20250139376A1 patent drawing

AI summary

Provided is an information processing system configured to: acquire a question based on input from a user; acquire answer basis information corresponding to the question from a database; request a language model to create an answer to the question based on the acquired answer basis information and acquire a model answer from the language model; determine whether the model answer matches the acquired answer basis information; and generate an answer to be output to the user based on a result of the determination and the model answer.