This repository contains the implementation of a modular, machine learning-based architecture for the automatic evaluation of scientific research papers, using Natural Language Processing (NLP) techniques and Large Language Models (LLMs). This work was done as part of my thesis, which is available in PDF form in this repository.
The system performs:
- Document parsing from PDF to raw text
- Section segmentation using a hybrid LSTM–BERT model
- Post-processing and correction of segmentation inconsistencies
- Section-level classification into accepted/rejected labels
- Document-level decision aggregation based on section predictions
| TASK | ACCURACY |
|---|---|
| Section Segmentation | 80% |
| Section-Level Classification | 77% |
| Paper-Level Decision (with CWA) | 92% |
Unfortunately due to github constraints the full training data and the weights of the models cannot be uploaded. However, if you are interested in obtaining either you can me send a message and we'll figure something out. That being said, training data for the classification part can be easily acquired through the scripts. For segmentation a lot of manual labeling was required so it's not as easy.