Projects

Automatic Title Generation From Abstracts Of Research Articles

Bachelor's Thesis | Areas covered: NLP, Deep Learning, Text Generation, Rule-Based Methods
December 2021 - July 2022
Tech Stack: Python, NumPy, Spacy, NLTK, Scikit-Learn, PyTorch, Simple Transformers, Matplotlib, Seaborn
Code    Project Report    Slides    Live Demo (Video)

• Constructed 5 datasets from 5 domains- AI, Mathematical Physics, Astrophysics, Cryptography, Neuroscience by scraping data from ArXiV - To test robustness of models across multiple domains
• Implemented a Dependency Tree based model (DTATG) from scratch based on a pre-existing paper - Model requires no training data
• Trained a bidirectional LSTM model with Attention Mechanism from scratch : Did not perform well
• Fine-tuned BART (Bidirectional Autoregressive Transformer) model : Performs well but gives generic titles
• Developed new model (Keyword-guided BART : KEG-BART) using fine-tuned BART & a module from DTATG to feed modified (smaller & information-dense) abstracts into BART for generating more specific titles
  The algorithm developed by us in this project (KEG-BART) is shown below: