Deep Natural Language Processing Project: Topic Modeling

This repository focuses on topic modeling techniques that leverage BERT-based keyword extraction. We explore three main approaches:

1. Domain Adaptation – Applying keyword extraction in a specialized domain (e.g., agriculture).
2. Multilingual Extension – Handling documents in the Greek language.
3. NER-based Preprocessing – Using Named Entity Recognition to filter key entities before extracting keywords.

Repository Structure

• domain_adaptation/ covers the agriculture-domain adaptation approach.
• multilingual/ includes all code for multilingual (Greek) modeling.
• ner_preprocessing/ implements NER-based entity filtering.
• utils/ has utility scripts for logging, helper functions, etc.

How to Run

1. Install Dependencies

   pip install spacy nltk scikit-learn requests
   pip install torch sentence-transformers keybert thefuzz
   python -m spacy download el_core_news_sm
   

2. Running via main.py

   We provide a single entry point in main.py that accepts a parameter specifying which approach to run:

   python main.py --approach domain

   Runs the domain adaptation pipeline.

   python main.py --approach multilingual

   Runs the multilingual (Greek) pipeline.

   python main.py --approach ner

   Runs the NER-based preprocessing pipeline.

   Inside main.py, these commands map to the corresponding scripts in their respective folders.

Results

During each run, the code may generate:

• Logs: Training and validation logs for model performance tracking.
• Metrics: Precision, Recall, F1 scores for keyword extraction.
• Comparison: We compare the final results (baseline vs. extended approaches) in our final report.