DocuClassif: A Document Classification Pipeline

Models:

• Zero-Shot Classification with Large Language Models (LLM)
• Support Vector Machine, using TD-IDF of n-grams as features.

Dataset used for Training/Testing:

• Multi-Eurlex and associated paper.

Use of API:

Launch FastAPI using uvicorn main:app with --reload if in testing mode.

Then, the API can be called by using the following instructions:

curl -X 'POST' 'http://127.0.0.1:8000/classification' \
    -H 'accept: application/json' \
    -H 'Content-Type: application/json' \
    -d '{"text":"Make decisions 10X more confidently and quickly with AI-powered insights.", \
         "labels":["Business","Machine Learning","Sport"], \
         "model": "zero-shot"}'

The ouput is a dict with the predicted labels. model is to be chosen in ['svm','zero-shot']. labels is an optional input if the zero-shot model is used (no specification leads to a default list specified in params.py). Note that SVM model only works with the EuraLex Labels (specified in params).

File Structure:

• main.py is the API structure.
• params.py contains the main Hyperparameter.
• test.py is an inner testing file to to compare performance of models.
• For the SVM Model,
  • svm_classif.py is the SVM classifier trained on Multi-Eurlex and stored in models/svm.
  • Note that svm_classif_training.py allows to train the SVM model on a given dataset (e.g. EuraLex) while performing a Grid Search for Hyperparameter Tunning.
• zero_classif.py is the Zero-Shot NLI Classifier imported from Hugging Face Library (default is facebook/bart-large-mnli).
• frontend.py allows to call a Gradio Frontend to interact with the API.

Some ideas of improvement:

• (ML) Few-shot Learning adding a last layer of prediction for the label class (i.e. classify embeddings).
• (ML) Mix SVM and Few/zero-shot learning for Long Document Purpose.
• (Soft. Eng.) Dockerize the Python Server, following for instance 1.
• Extend to [Lex Glue Dataset](LexGLUE: A Benchmark Dataset for Legal Language Understanding in English).

Credit to:

• LexGlue
• lmtc-eurlex57k
• muit-eurlex