Classification Model Training With Unlabeled AI Short-Text Samples

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional AI-generated text detection methods struggle with accuracy when the input text is short, as the length of the text affects detection difficulty, making it challenging to distinguish between human-generated and AI-generated content accurately.

Innovation Solution

The method involves generating target training data with a smaller data amount based on original training data, marking AI-generated data as unlabeled, and using a loss function that considers positive and unlabeled samples to improve the classification model's focus on AI-generated data, enhancing prediction accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If conventional binary classification is used for AI-generated text detection, then the detection process is simple, but prediction accuracy deteriorates when input text is short

Engineering Contradiction:
Improvedetection process complexityVSAvoidprediction accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent segments the training data into multiple groups based on text length, with each group containing both positive samples (user-generated) and negative samples (AI-generated). This segmentation allows the model to learn from balanced data across different length scales, improving prediction accuracy for short texts while maintaining a manageable detection process through systematic data organization.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter of text length by creating multiple training data groups with different length ranges. Each group is specifically designed to handle texts of certain length characteristics, allowing the model to adapt its prediction behavior based on the input text length, thereby improving accuracy for short texts without overly complicating the overall detection framework.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If training data is divided into multiple groups by text length, then prediction accuracy for short texts is improved, but data processing complexity increases

Engineering Contradiction:
Improveprediction accuracyVSAvoiddata processing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by pre-dividing the training data into multiple groups based on text length before model training. This preprocessing step organizes the data into structured groups with balanced positive and negative samples, making the subsequent training process more efficient. The preliminary organization reduces the complexity during actual model execution while maintaining high prediction accuracy across different text lengths.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If AI-generated short texts are marked as unlabeled, then the model focuses better on AI-generated data characteristics, but training complexity increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidtraining complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent introduces an intermediary approach by marking AI-generated short texts as unlabeled samples rather than directly labeling them. This creates a semi-supervised learning scenario where the model learns to identify AI-generated characteristics through the unlabeled data while being guided by the labeled positive samples. This intermediary strategy improves the model's ability to detect AI-generated short texts without requiring complex retraining procedures.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4685673A1Method for training classification model, and related apparatus
Publication Date: 2026.01.28 HUAWEI TECH CO LTD
  • EP4685673A1 patent drawingFigure 1
  • EP4685673A1 patent drawingFigure 2
  • EP4685673A1 patent drawingFigure 3

AI summary

A classification model training method is provided, and is applied to the field of artificial intelligence technologies. In the method, target training data with a smaller data amount is first generated based on original training data, to obtain training data of different scales, and ensure scale diversity of the training data. In addition, considering a characteristic of AI-generated data, for target training data that has a small data amount and that is originally AI-generated, this part of target training data is marked as unlabeled data, so that during training, a classification model can focus on data that is AI-generated and that is different from manually generated data, to avoid affecting perception of an AI-generated data style by the classification model when the data is marked as the AI-generated data, and effectively improve prediction accuracy of the classification model.