Document Metadata Extraction with LLM
Automatic extraction of bibliographic metadata from academic documents (theses, articles, books, conference objects) using fine-tuned language models. Built around the SEDICI repository (Servicio de Difusion de la Creacion Intelectual) at Universidad Nacional de La Plata.
Project Goal
Given an academic PDF document, automatically extract structured metadata such as title, authors, date, language, rights, and type-specific fields (director, ISBN, ISSN, etc.) using fine-tuned LLMs and ML classifiers.
High-Level Flow
flowchart LR
A[SEDICI\nRepository] -->|Download & Clean| B[Dataset\n2000 docs]
B -->|Train| C[Fine-Tuned\nLLM]
B -->|Train| D[Type & Subject\nClassifiers]
C --> E[API\nMicroservices]
D --> E
E -->|Upload PDF| F[Structured\nMetadata JSON]Project Structure
| Folder / File | Description |
|---|---|
api/ |
FastAPI microservices (Orchestrator, Extractor, LLM Service) |
download_prepare_clean_normalize_sedici_dataset/ |
Data preparation pipeline: download, extract text, clean metadata |
fine_tunning/ |
LLM fine-tuning module (LED, LLAMA, GEMMA, etc.) |
fine_tune_type/ |
Document type classifier (Libro, Tesis, Articulo, Objeto de conferencia) |
fine_tune_subject/ |
Subject/topic classifier using multiple ML models (SVM, XGBoost, etc.) |
validation/ |
Model evaluation and comparison against ground truth |
utils/ |
Shared utilities (text extraction, normalization, API clients) |
data/ |
Data storage (CSVs, PDFs, extracted texts, JSONs) |
constants.py |
Global configuration, prompts, model definitions, field mappings |
run_modules.sh / run_modules.bat |
Entry-point scripts to run each module |
Supported Document Types
| Type | Spanish | Example Specific Fields |
|---|---|---|
| Thesis | Tesis | director, codirector, degree |
| Book | Libro | publisher, ISBN |
| Article | Articulo | ISSN, journal |
| Conference Object | Objeto de conferencia | ISSN, event |
Technologies
- Deep Learning: PyTorch, Transformers (Hugging Face), PEFT, LoRA
- Base Models: LED, LLAMA, GEMMA, Mistral, T5, NuExtract
- Document Processing: pdfplumber, PyMuPDF, EasyOCR
- ML Classifiers: scikit-learn (SVM), XGBoost, Random Forest
- API: FastAPI, Docker, Docker Compose
- Data Cleaning: Google Gemini API, OpenAI API (configurable)