Language Model Training with Reliable Source Filtering

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional information systems lack reliability in the generation of answers due to varying degrees of reliability in news article sources.

Innovation Solution

An information providing apparatus that collects data from reliable sources, learns a language model using high-reliability data, and generates answers tailored to user preferences, adjusting learning based on user feedback and service plans.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If data is collected from any provider, then the quantity of data is increased, but the reliability of the information decreases

Engineering Contradiction:

Improvequantity of dataVSAvoidreliability of information

Core Design Contradiction:

Quantity of substanceVSReliability

Solution Approach 1:

The system segments providers into different reliability groups based on evaluation results, and selectively collects data only from high-reliability providers. This segmentation allows the system to maintain data quantity while ensuring quality by dividing the provider base into manageable reliability tiers and prioritizing those that meet specified criteria.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system changes the parameter of provider selection by introducing reliability evaluation metrics and dynamically adjusting which providers are selected for data collection. By evaluating providers against changing reliability parameters and selecting based on these evaluations, the system ensures that data quantity increases only from sources that meet the required reliability threshold.

Inventive Principle:

Principle #35Parameter changes

2Reliability

If data is collected only from reliable providers, then the reliability of information is improved, but the quantity of data is reduced

Engineering Contradiction:

Improvereliability of informationVSAvoidquantity of data

Core Design Contradiction:

ReliabilityVSQuantity of substance

Solution Approach 1:

The system performs preliminary evaluation of providers before data collection to identify and select only those meeting reliability criteria. By conducting this preliminary assessment and selection process in advance, the system ensures that subsequent data collection from chosen providers maximizes both reliability and available data quantity within the constrained provider base.

Inventive Principle:

Principle #10Preliminary action

3Productivity

If a language model is generated using all available data, then the productivity of model generation is increased, but the reliability of generated answers decreases

Engineering Contradiction:

Improveproductivity of model generationVSAvoidreliability of generated answers

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The system extracts and uses only the high-reliability portion of available data for language model generation, separating reliable data from unreliable data. By taking out and utilizing only the subset of data from evaluated high-reliability providers, the system maintains efficient model generation productivity while ensuring that the training data quality improves the reliability of generated answers.

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentEP4715674A1Information provision device, information provision method, and information provision program

Publication Date: 2026.03.25 SOFTBANK GROUP CORP

Figure 1
Figure 2~3
Figure 4~5

AI summary

An information providing apparatus according to an embodiment includes a collection unit and a generation unit. The collection unit collects a data source provided from a provider satisfying a condition regarding reliability. The generation unit generates a language model generated by learning the collected data source and configured to generate an answer to a prompt input by a user.