Rishabbh-Sahu / ignore_email_address_classifier

Important vs Ignore email classifier based on incoming email addresses. Bert tokenizer is used as a tokenization method and CNN network as the model. Framework used - Tensorflow 2.4

ignore_email_address_classifier

About this project

This repo can be used for any text (english lanugage) classification task leveraging bert-tokenizer, CNN-layers and "Tensorflow 2.4" framework. By using a few specific keyworks like "donotreply", "no-reply", "unsubscribe" etc. in the email address, I'd created labels as important/ignore-emails to enable supervised learning. Extended the vary concept of tokenization to email addresses and split them into sub-tokens (a.k.a. sub-words). Using the bert sub-tokens embedding and training them using CNN architecture, achieved very good accuracy in a very few epochs. To enable GPU support, please do enable CUDA-11 in windows/linux/mac virtual environment for tf2.4 or use CUDA 10.1 for tf2.3.

Getting started

• create virtual environment
• install tensorflow==2.4
• install requirements
• Open config.yaml file and modify parameters as per your setup. This will also allow us to perform experiments and refer them back for comparison.

Data source

https://www.kaggle.com/wcukierski/enron-email-dataset - Post reviewing some random emails, created the label (important vs ignore) based on email-ids. for e.g. some keywords like noreply, unsubscribe, information etc. relates to marketing or emails from unknown sources at times hence falls under the category of ignore-emails.

For training

• python training.py

Model validation

• Training accuracy after 1-epoch ~99%
• Validation accuracy ~99%
  Note: Try running for more epochs if you have very less data to get good results

Experiments (modify config.yaml accordingly):

1. Instead of using bert tokenizer, we can create our own word embeddings based on the dataset by using embedding layer and tune embedding's dimension based on f1-score OR used pre-learned ones like word2vec, GloVe etc. - for the reference, pls do visit model.py (updated links)
2. BPE (byte pair encoding) can be explored for text tokenization with vocab size 2k, 5k, 10k etc. Use this link to explore it further - https://github.com/rsennrich/subword-nmt/blob/master/subword_nmt/learn_bpe.py

Other use-cases:

To perform any text classification tasks like intent, sentiments, toxic comments identification etc. Below are some open source datasets for the reference.

Future scope:

Add manual intervention to input the email-id for classification.

Datasets to explore:

• Toxic-comment-classification: https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/data
  
• Sentiment analysis of IMDB reviews: https://www.kaggle.com/lakshmi25npathi/imdb-dataset-of-50k-movie-reviews

Reference:

https://stackabuse.com/text-classification-with-bert-tokenizer-and-tf-2-0-in-python/