Talking Stage v2.0

What's New in TalkingStageBot v2.0

We are excited to announce the release of TalkingStageBot v2.0! This new version comes with significant enhancements and improvements to make the chatbot smarter and more efficient. Here’s what’s new:

Key Features and Improvements

1. Enhanced Machine Learning Model

• Upgraded ML Model: The machine learning model has been upgraded to use the OneVersusAll trainer with the LbfgsLogisticRegression binary classifier. This provides better accuracy and performance for multiclass classification.
• Regularization Parameters: Fine-tuned L1 and L2 regularization parameters for optimal model performance.
  • L1 Regularization: 0.031347524F
  • L2 Regularization: 0.03125F

2. Improved Data Preprocessing

• Text Normalization: Text data is now preprocessed to ensure all text is in lowercase and trimmed, which helps in reducing the complexity and improving the model's ability to generalize.
• Data Pipeline: Enhanced data pipeline includes text featurization, feature concatenation, and key-to-value mapping for labels.

3. Robust Error Handling

• Platform Compatibility: Added specific handling for PlatformNotSupportedException to ensure smooth operation across different platforms.
• Type Initialization: Improved error handling for type initialization issues to make debugging easier and enhance stability.

4. Multi-Platform Support

• Cross-Platform Compatibility: TalkingStageBot is now compatible with multiple platforms, including macOS and Windows.

5. Simplified Training Process

• Streamlined Training and Saving: The process of training and saving the model has been streamlined with comprehensive debug logs to track the progress and catch errors effectively.

Talking Stage v1.3.0

What's New in v1.3.0

Key Updates

Enhanced User Interaction

• New Input Method: Users can now send their messages to the bot by pressing the enter key or clicking the send button. This makes the interaction smoother and more intuitive.

Improved Machine Learning Model

• Algorithm Parameter Tuning: Updated the machine learning algorithm parameters for better performance. The LightGBM model now has the following configuration:
  • Number of Leaves: 20
  • Number of Iterations: 100
  • Minimum Example Count Per Leaf: 5
  • Learning Rate: 0.05
  • Gradient Booster Options:
    • Subsample Fraction: 0.8
    • Feature Fraction: 0.8
    • L1 Regularization: 0.1
    • L2 Regularization: 0.2
  • Maximum Bin Count Per Feature: 254
  • Early Stopping Rounds: 20

Expanded Training Data

• Increased Data Points: The training dataset has been expanded to over 3500 data points. This larger dataset helps in making the model more robust and accurate.

Model Pipeline Enhancements

• Preprocessing Steps: The training data is now preprocessed to ensure consistency. This involves trimming and converting text to lowercase to avoid discrepancies.
• Feature Engineering: Enhanced feature engineering steps in the training pipeline to improve model performance.

Talking Stage v1.2.3

What's New in v1.2.3 Release

• Implemented a more advanced machine learning algorithm (LightGBM) for improved prediction accuracy.
• Fine-tuned model parameters for enhanced performance and reliability.
• Optimized data preprocessing steps to ensure consistent and effective training.
• Updated data loading and preprocessing to handle text data more efficiently.
• Streamlined the training pipeline for faster model training and deployment.
• Enhanced model robustness with increased iterations and improved feature handling.
• Integrated comprehensive error handling and logging for better debugging capabilities.
• Improved overall code readability and maintainability.

Talking Stage v1.2.1

Changelog for v1.2.1

[v1.2.0] - 2024-06-25

Added

• Initial Interaction Alert: On the first interaction, the bot now displays an alert with guidance on how to format questions for better responses. This helps users understand the input format for optimal results.
• Unique Responses: Updated the response mechanism to ensure that each response is unique and separated by a comma.
• Expanded Training Data: The training data has been expanded to better handle a wider variety of questions, resulting in more accurate and relevant answers.

Changed

• Response Method Update: The GetResponse method now returns a tuple containing the response and a flag indicating whether it is the first interaction.
• MainPage Interaction Handling: Modified the OnSendClicked method in MainPage.xaml.cs to handle the first interaction alert display based on the flag from GetResponse.

Fixed

• Improved Keyword Matching: Enhanced keyword matching in responses to ensure more accurate identification of user queries.

How to Update

To update your local repository to this version, use the following git commands:

git fetch origin
git checkout v1.2.0

How to Use

1. On the first interaction, the bot will display an alert with the message: "Note: Different questions are identified and answered when separated by a comma, question mark, full stop or exclamation marks."
2. Input multiple questions separated by punctuation marks for the bot to provide distinct, relevant responses.

Known Issues

• The bot may still provide "I don't have an answer for that." for queries that are out of the scope of the training data.

Future Plans

• Further expansion of the training dataset to cover more topics.
• Continuous improvement of the machine learning model for better accuracy and relevancy in responses.

Talking Stage v1.0.2

Changes Made:

Implemented model saving and loading functionality.
Refactored code to separate model training from bot functionality.
Updated README.md to document new features, setup instructions, and development guidelines.

• Additional Notes:
  This version (v1.0.2) represents a significant improvement over the pre-release version by enhancing performance and user experience through optimized model handling and clearer documentation.

First Release

This first release is not production ready, because it doesnt create a prediction model file saved into the project folder, instead it creates a prediction engine from the training data through high level abstractions provided by ML.NET framework, hence simplifying it.

Subsequent release would seperate the creation of the prediction model file from the live prediction itself by checking if the prediction model file is present first and then loading it to be used directly for prediction using Transform directly on the IDataView.