TherapeuticNLP 1.0.0

A multi-dimensional Natural Language Processing (NLP) framework for analyzing and assessing the quality of therapeutic conversations

Therapeutic Conversation Quality Assessment Framework

A multi-dimensional Natural Language Processing (NLP) framework for analyzing and assessing the quality of therapeutic conversations.

🌟 Overview

This project presents a novel approach to evaluating therapeutic conversation quality using advanced NLP techniques and machine learning. Our framework analyzes key conversational dynamics to distinguish between high and low-quality therapeutic interactions, providing valuable insights for mental health professionals.

🔑 Key Features

• Multi-dimensional Analysis: Evaluates conversations across four key dimensions:

  • Conversation Analytics (turn-taking, word usage patterns)
  • Semantic Analysis (topic coherence and flow)
  • Sentiment Analysis (emotional context)
  • Question Detection (engagement patterns)

• Advanced ML Classification: Implements multiple classifiers including Random Forest, CatBoost, and SVM, achieving up to 97% accuracy with optimized parameters

• Robust Data Processing:

  • Handles imbalanced datasets using SMOTE-Tomek
  • Comprehensive outlier detection
  • Feature normalization and preprocessing

System Arhcitecture

📊 Performance Highlights

Classifier	Accuracy	Precision	Recall	F1 Score	AUC Score
SVM	0.9717	0.9775	0.9667	0.9715	0.9874
CatBoost	0.9600	0.9606	0.9600	0.9600	0.9912
Random Forest	0.9533	0.9487	0.9600	0.9539	0.9893

🛠 Technical Implementation

Feature Extraction Pipeline

1. Conversation Analytics:

   • Words per turn analysis
   • Turn-taking patterns
   • Statistical measures (std dev, skewness, kurtosis)

2. Semantic Analysis:

   • Utilizes multiple embedding models:
     • PromCSE
     • Sentence-BERT
     • SAKIL sentence similarity
   • Both overall and turn-order-aware analysis

3. Sentiment Analysis:

   • Twitter-roBERTa-base model
   • Sentiment transition tracking
   • Weighted certainty scores

4. Question Detection:

   • Syntactic pattern recognition
   • Bi-gram analysis
   • Speaker-specific question tracking

📦 Requirements

• Python 3.8+
• scikit-learn
• transformers
• torch
• pandas
• numpy
• catboost
• SVM
• Random Forest

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

This research was supported by:

• Natural Sciences and Engineering Research Council of Canada (NSERC)
• New Frontiers in Research Fund
• LeaCros

📬 Contact

For questions and feedback, please contact Niloy Roy