HNATT

This is a Keras implementation of the Hierarchical Network with Attention architecture (Yang et al, 2016), and comes with a webapp for easily interacting with the trained models.

Overview

HNATT is a deep neural network for document classification. It learns hierarchical hidden representations of documents at word, sentence, and document levels. At both the word and sentence levels, HNATT makes use of an attention mechanism, in which it learns a context vector that determines a relevance weighting for its learned encoding of words and sentences. This model has been shown to outperform hierarchical models without attention, indicating that learning an input-dependent weighting across the various substructures of the document leads to improved performance in classification tasks.

Contents

Module Description
hnatt.py* Main HNATT implementation with custom Attention layer.
util.yelp.py Data loader for Yelp review data used for training and testing.
util.text_util.py Utility function for normalizing review texts.
util.glove.py Utility function for loading GloVe embedding weights from a file. You can download the embeddings here.
main.py Demo that trains HNATT on a subset of Yelp reviews and displays attention activation maps at both sentence and word levels on an example review.
app/ A simple Flask app for exploring a trained HNATT, allowing you to easily make predictions based on a text input and visualize the resulting attention activations at both word and sentence levels.

*A TensorFlow backend is assumed by the Attention layer.

Get started

Install dependencies in a new virtual environement via

virtualenv .venv
source .venv/bin/activate
pip install -r requirements.txt

Download the latest Yelp dataset from https://www.kaggle.com/yelp-dataset/yelp-dataset and move the unzipped folder to data/yelp-dataset in the project directory.

Give it a spin.

python main.py

Train your model

First, load n reviews from yelp for training, with 90/10 training/test split:

import util.yelp as yelp
(train_x, train_y), (test_x, test_y) = yelp.load_data(path=YELP_DATA_PATH, size=1e5, train_ratio=0.9)

You can also choose to polarize the Yelp ratings into binary labels, corresponding to negative and positive sentiment, by passing in an optional argument, binary=true. The resulting training and test sets will be balanced.

Now you're ready to train your model:

from hnatt import HNATT
h = HNATT()	
h.train(train_x, train_y, 
	batch_size=16,
	epochs=16,
	embeddings_path=EMBEDDINGS_PATH, 
	saved_model_dir=SAVED_MODEL_DIR,
	saved_model_filename=SAVED_MODEL_FILENAME)

You can print out sentence and word-level attention activations like so:

activation_maps = h.activation_maps('loved it! will definitely go back again.')
print(activation_maps)

Performance

When trained on random samples of 10,000 Yelp reviews, loaded with binary=true and balanced classes in training and test sets, this implementation of HNATT reaches 100% accuracy on the test set, and consistently around 90% accuracy on the validation set.

Visualizing attention

Once you train an HNATT model and save it locally using the saved_model_dir and saved_model_filename arguments to train, you can easily play with the saved model in an interactive web app by running the following:

python run_hnatt_viewer.py

You can then visit localhost:5000 to interact with your HNATT.

GitHub