Text Summarization models

This repo is built to collect multiple implementations for abstractive approaches to address text summarization , for different languages (Arabic , English)

it is built to simply run on google colab , in one notebook so you would only need an internet connection to run these examples without the need to have a powerful machine , so all the code examples would be in a jupiter format , and you don't have to download data to your device as we connect these jupiter notebooks to google drive

• Arabic Summarization Model using the corner stone implemtnation (seq2seq using Bidirecional LSTM Encoder and attention in the decoder) for summarizing Arabic news
• implementation A Corner stone seq2seq with attention (using bidirectional ltsm ) , three different models for this implemntation
• implementation B seq2seq with pointer genrator model
• implementation C seq2seq with reinforcement learning

Blogs

This repo has been explained in a series of Blogs
- to understand how to work with google colab eco system , and how to integrate it with your google drive , this blog can prove useful DeepLearning Free Ecosystem
- Tutorial 1 Overview on the different appraches used for abstractive text summarization
- Tutorial 2 How to represent text for our text summarization task
- Tutorial 3 What seq2seq and why do we use it in text summarization
- Tutorial 4 Multilayer Bidirectional Lstm/Gru for text summarization
- Tutorial 5 Beam Search & Attention for text summarization
- Tutorial 6 Build an Abstractive Text Summarizer in 94 Lines of Tensorflow
- Tutorial 7 Pointer generator for combination of Abstractive & Extractive methods for Text Summarization

Try out this text summarization through this website (eazymind) ,
which enables you to summarize your text through
- curl call

curl -X POST 
http://eazymind.herokuapp.com/arabic_sum/eazysum
-H 'cache-control: no-cache' 
-H 'content-type: application/x-www-form-urlencoded' 
-d "eazykey={eazymind api key}&sentence={your sentence to be summarized}"

• python package (pip install eazymind)
  pip install eazymind

from nlp.eazysum import Summarizer

#---key from eazymind website---
key = "xxxxxxxxxxxxxxxxxxxxx"

#---sentence to be summarized---
sentence = """(CNN)The White House has instructed former
    White House Counsel Don McGahn not to comply with a subpoena
    for documents from House Judiciary Chairman Jerry Nadler, 
    teeing up the latest in a series of escalating oversight 
    showdowns between the Trump administration and congressional Democrats."""

summarizer = Summarizer(key)
print(summarizer.run(sentence))

Implementation A (seq2seq with attention and feature rich representation)

contains 3 different models that implements the concept of hving a seq2seq network with attention
also adding concepts like having a feature rich word representation
This work is a continuation of these amazing repos

Model 1

is a modification on of David Currie's https://github.com/Currie32/Text-Summarization-with-Amazon-Reviews seq2seq

Model 2

1- Model_2/Model_2.ipynb

a modification to https://github.com/dongjun-Lee/text-summarization-tensorflow

2- Model_2/Model 2 features(tf-idf , pos tags).ipynb

a modification to Model 2.ipynb by using concepts from http://www.aclweb.org/anthology/K16-1028

Results

A folder contains the results of both the 2 models , from validation text samples
in a zaksum format , which is combining all of
- bleu
- rouge_1
- rouge_2
- rouge_L
- rouge_be
for each sentence , and average of all of them

Model 3

a modification to https://github.com/theamrzaki/text_summurization_abstractive_methods/blob/master/Model_3.ipynb

Implementation B (Pointer Generator seq2seq network)

it is a continuation of the amazing work of
https://github.com/abisee/pointer-generator
https://arxiv.org/abs/1704.04368
this implementation uses the concept of having a pointer generator network to diminish some problems that appears with the normal
seq2seq network

Model_4_generator_.ipynb

uses a pointer generator with seq2seq with attention
it is built using python2.7

zaksum_eval.ipynb

built by python3 for evaluation

Results/Pointer Generator

• output from generator (article / reference / summary) used as input to the zaksum_eval.ipynb
• result from zaksum_eval

i will still work on their implementation of coverage mechanism , so much work is yet to come if God wills it isA

Implementation C (Reinforcement Learning For Sequence to Sequence )

this implementation is a continuation of the amazing work done by
https://github.com/yaserkl/RLSeq2Seq
https://arxiv.org/abs/1805.09461

@article{keneshloo2018deep,
 title={Deep Reinforcement Learning For Sequence to Sequence Models},
 author={Keneshloo, Yaser and Shi, Tian and Ramakrishnan, Naren and Reddy, Chandan K.},
 journal={arXiv preprint arXiv:1805.09461},
 year={2018}
}

Model 5 RL

this is a library for building multiple approaches using Reinforcement Learning with seq2seq , i have gathered their code to run in a jupiter notebook , and to access google drive
built for python 2.7

zaksum_eval.ipynb

built by python3 for evaluation

Results/Reinforcement Learning

• output from Model 5 RL used as input to the zaksum_eval.ipynb