Distinguishing Dementia Using Hebrew NLP and LLM Models

This project focuses on developing a machine learning pipeline to distinguish early-stage dementia among Hebrew-speaking individuals. Using Natural Language Processing (NLP) techniques and Large Language Models (LLMs), the goal is to analyze speech patterns and identify linguistic markers associated with dementia. The approach incorporates a distribution and variety of speech characteristics from diverse individuals.

---

Key Objectives

1. Data Collection: Gather a diverse dataset of speech samples from Hebrew-speaking individuals, including those diagnosed with dementia and healthy controls.
2. Feature Extraction: Use NLP techniques to extract linguistic features such as syntax, semantics, fluency, and speech dynamics.
3. Model Development: Fine-tune Hebrew-compatible LLMs to classify dementia-related speech patterns.
4. Evaluation: Measure model accuracy and generate insights into linguistic differences between healthy individuals and those with dementia.

---

Repository Structure

---

Steps in the Process

1. Data Collection:

   • Collect speech samples from clinical datasets (dementia patients) and healthy individuals.
   • Convert audio files to text using Hebrew-compatible speech-to-text systems.

2. Data Preprocessing:

   • Clean and normalize Hebrew text for NLP tasks.
   • Tokenize, lemmatize, and tag parts of speech using Hebrew NLP tools.

3. Feature Engineering:

   • Extract linguistic features:
     • Lexical Features: Vocabulary richness, word frequency.
     • Syntactic Features: Grammar patterns and sentence structure.
     • Semantic Features: Contextual coherence and sentiment analysis.
     • Prosodic Features: Speech rate and intonation (if audio is used).

4. Model Development:

   • Fine-tune Hebrew-compatible LLMs:
     • AlephBERT, HeBERT, or multilingual models like mBERT and XLM-RoBERTa.
   • Train models to classify dementia-related speech patterns.

5. Model Evaluation:

   • Evaluate models using metrics like Accuracy, Precision, Recall, and F1-score.
   • Use t-SNE or PCA to visualize differences between dementia and non-dementia speech.

---

Tools and Technologies

• Programming: Python
• NLP Libraries: transformers, spaCy, nltk
• Speech-to-Text: Google Speech API, Wav2Vec2 for Hebrew
• Machine Learning: TensorFlow, PyTorch, scikit-learn
• Visualization: Matplotlib, Seaborn

---

How to Run the Project

1. Clone the repository:

   git clone https://github.com/your-username/Dementia-NLP-Hebrew.git