Train Classification Models

TextGO can train machine learning models to recognize custom text types.

What Is a Classification Model?

TextGO uses TensorFlow.js to train and run models locally in the app's WebView:

• No Backend Required: Training and inference run locally
• Privacy and Security: Data never leaves your device
• Real-time Inference: Loaded models recognize text quickly
• Lightweight: Models are small and load quickly

When to Use Classification Models

Suitable Scenarios for Models

✅ Complex Text Patterns

• Patterns that are difficult to describe with regular expressions
• Text with variations but consistent overall characteristics

✅ Sufficient Training Data

• At least 3 unique, non-empty positive samples; more diverse samples generally improve results
• Samples should cover main variations

✅ Some Misclassification Is Acceptable

• Workflows can tolerate occasional misclassification
• Suitable for fuzzy matching

Unsuitable Scenarios for Models

❌ Simple and Precise Patterns

• Patterns that regular expressions can describe precisely
• Phone numbers, ID numbers, and other fixed formats

❌ Insufficient Training Data

• Fewer than 3 valid unique samples cannot be used for training
• Samples do not cover the main variations

❌ Misclassification Is Unacceptable

• Workflows with strict accuracy requirements

Create a Classification Model

Step 1: Access Model Management

1. Open "Settings" > "Classification Model"
2. Click the "+" button to open the "New Classification Model" dialog

Step 2: Basic Information

Fill in the model's basic information:

Type Name (Required)

• Identifies the model
• Use a descriptive name

Type Icon (Optional)

• Click the current type icon to open the icon selector
• Select from "Built-in Icons" or use "Upload Custom SVG"

Step 3: Prepare Training Data

Positive Samples (Required)

Positive samples are key to determining the model's recognition capability.

Data Format:

• Enter one sample per line
• Samples can contain any text
• Blank lines are ignored and duplicate samples are removed

Sample Quality Requirements:

• ✅ Samples should cover main text variations
• ✅ Samples should include typical characteristics of this text type
• ✅ Remove irrelevant content from samples
• ❌ Avoid providing identical samples
• ❌ Avoid including erroneous or invalid samples

Step 4: Configure Parameters

Basic Parameters

Confidence Threshold (Required, 1%–99%)

• Default: 50%
• Text with similarity greater than or equal to this threshold will be recognized as this type
• Adjust as needed:
  • Increase threshold → Stricter matching, reduces false positives
  • Decrease threshold → Looser matching, increases recognition rate

Use a Classification Model

Trained models appear in the recognition type list:

1. Open "Global Shortcuts"
2. Add a new rule
3. Select the trained model in "Recognize Type"
4. Configure an action and save