This repository contains the backend microservices responsible for the core AI-powered analysis and PDF processing functionality of the IMRaD Introduction Analysis platform. This platform is developed as part of my graduation thesis.
Parent Repository (Next.js Frontend & API): https://github.com/stormsidali2001/graduation_IMRAD_introduction_analysis_SaaS
This repository hosts two independent microservices:
-
AI Models Microservice (FastAPI):
- Handles interaction with the TensorFlow Serving service to get predictions from the trained AI models.
- Provides additional features like summarization and author thought process generation, utilizing the Gemini API.
-
PDF Extractor Microservice (FastAPI):
- Extracts the introduction text from research papers uploaded in PDF format.
The platform leverages four fine-tuned BERT models hosted on Hugging Face:
- IMRaD Introduction Move Classifier: https://huggingface.co/stormsidali2001/IMRAD_introduction_moves_classifier
- IMRaD Introduction Move 0 Sub-move Classifier: https://huggingface.co/stormsidali2001/IMRAD-introduction-move-zero-sub-moves-classifier
- IMRaD Introduction Move 1 Sub-move Classifier: https://huggingface.co/stormsidali2001/IMRAD-introduction-move-one-sub-moves-classifier
- IMRaD Introduction Move 2 Sub-move Classifier: https://huggingface.co/stormsidali2001/IMRAD-introduction-move-two-sub-moves-classifier
These models are responsible for classifying sentences in introductions into their respective IMRaD moves and sub-moves.
The AI models were trained on a unique dataset of over 169,000 sentences. This dataset was generated using Google's Gemini Pro model and a custom-designed pipeline applied to a set of randomly selected introductions from the unarXive IMRaD Classification Dataset (Hugging Face): https://huggingface.co/datasets/saier/unarXive_imrad_clf.
The process of creating the dataset is documented in the /notebooks/v3 directory of the main (Next.js) repository: https://github.com/stormsidali2001/graduation_IMRAD_introduction_analysis_SaaS
TensorFlow Serving is used to efficiently serve the AI models for prediction. The Docker Compose configuration for TensorFlow Serving is located in this repository, in the directory for the AI Models Microservice.
/moves: Contains the code for the AI Models Microservice (FastAPI)./pdf-extractor: Contains the code for the PDF Extractor Microservice (FastAPI)./tensorflow-models: Contains the Docker Compose configuration for TensorFlow Serving.
- Python 3.9+
- FastAPI
- TensorFlow
- Docker
later
MIT