Text Normalization using Memory Augmented Neural Networks

The Text Normalization Demo notebook and the accompanying paper "Text Normalization using Memory Augmented Neural Networks" demonstrates an accuracy of 99.5% on the Text Normalization Challenge by Richard Sproat and Navdeep Jaitly. An earlier version of the approach used here has secured the 6th position in the Kaggle Russian Text Normalization Challenge by Google's Text Normalization Research Group.

Go straight to the Text Normalization Demo Notebook

Architecture

Two models are used for the purpose of text normalization. A XGBoost boost model first classifies a token as to-be-normalized or remain-self. The to-be-normalized tokens are then fed character-by-character to our proposed Sequence to Sequence DNC model.

More details about the architecture and implementation can be found in the original paper.

Results :

1. Normalizing English Data

2. Normalizing Russian Data

Semiotic Classwise Accuracy

How to run?

Requirements:

• Jupyter Notebook
• Anaconda Package Manager
• rest will be installed by anaconda (see below)

Follow these steps for a demonstration:

1. Clone the repo
2. Download and extract the required data.

$ sh setup.sh

2. Create & activate an environment using the provided file

$ conda env create -f environment.yml
$ source activate deep-tf

3. Start a Jupyter Notebook server
4. Open 'notebooks/Text Normalization Demo.ipynb'
5. Set the language to English or Russian below the 'Global Config' cell

lang = 'english'
# lang = 'russian'

6. Run the notebook

Full Requirements:

• numpy 1.13.3
• pandas 0.21.0
• matplotlib 2.1.0
• watermark 1.5.0
• seaborn 0.8.1
• sklearn 0.19.1
• xgboost 0.6
• tensorflow 1.3.0

Authors

1. Subhojeet Pramanik (http://github.com/subho406)
2. Aman Hussain (https://github.com/AmanDaVinci)

Acknowledgements

Differentiable Neural Computer, Tensorflow Implementation: https://github.com/deepmind/dnc