paulrinckens/han_for_doc_classification

Hierarchical Attention Networks for Document Classification

Hierarchical Attention Networks for Document Classification

An implementation of Hierarchical Attention Networks for Document Classification from the following paper:

Content

This repo contains the following:

• Implementation of the Network Architecture with the following technologies
  • Keras
  • Preprocessing with spaCy
  • Embedding Layer with pretrained Glove vectors
  • Tokenizing with Keras tokenizer
• Utilities to train the model on the Ten Thousand German News Article Dataset (https://github.com/tblock/10kGNAD)
• Jupyter Notebook used during initial development of the network
• REST API server developed with FastAPI for predicting on a trained model and display of the sentence and word attentions in html

Get started

• Download german word vectors pretrained by Deepset (https://deepset.ai/german-word-embeddings) from here https://int-emb-glove-de-wiki.s3.eu-central-1.amazonaws.com/vectors.txt and place them in the directory embeddings/glove_german/
• The dataset can be downloaded from https://github.com/tblock/10kGNAD/blob/master/articles.csv. Place the articles.csv file in the directory data/

Train the model

Run the training script train_script.py

Start the server

Running

uvicorn app.app:app

starts the server under http://127.0.0.1:8000

Access the API docs

After running the server the API docs can be accessed via http://127.0.0.1:8000/docs

Make predictions

Send a HTTP request to the running server, e.g.

curl --location --request GET 'http://127.0.0.1:8000/predict/?text=<your-text-to-be-classifier>'

Visualize prediction and attentions

Send a HTTP request to the running server, e.g.

http://127.0.0.1:8000/visualize/?text=<your-text-to-be-classifier>

The server responds with a static html which should look something like the following: