Text-Summarization Using Deep Learning

Jupyter notebooks for text summarization using Deep Learning techniques

-- Project Status: Active

Introduction

The purpose of this project is to produce a model for Abstractive Text Summarization, starting with the RNN encoder-decoder as the baseline model. From there, we come across the effectiveness of different methods for attention in abstractive summarization. These methods try to first understand the text and then rephrase it in a shorter manner, using possibly different words. For perfect abstractive summary, the model has to first truly understand the document and then try to express that understanding in short possibly using new words and phrases. We have used the concept of an encoder-decoder recurrent neural network with LSTM units and attention to generate summary from a given text.

Methods Used

• Word Embeddings using GloVe (Global Vectors)
• Encoder-decoder using RNN(Recurrent Neural Network)

Technologies

• Python
• Keras Library
• TensorFlow
• Jupyter
• etc.

Description

In this project we have used a sample dataset of news articles (CNN , Daily Mail). Currently we are facing a problem in implementing the pointer-generator network.

Architecture

Learning Rate Configuration

CyclicLR(mode='triangular2', base_lr= 0.2, max_lr= 0.001, step_size= (len(padded_sorted_texts)*0.9/BATCH_SIZE) * 2)

Word Embeddings

• ConceptNet Numbernatch word embeddings were used to encode the word meanings

Getting Started

1. Clone this repo (for help see this tutorial).

2. Raw Data is being kept on the local storage at the location ~/Text-Summarization/Original_data/cnn/stories

3. Data processing/transformation scripts are being kept [here](Repo folder containing data processing scripts/notebooks)

4. Installation steps: Use single backticks to call out code or a command within a sentence.

To format code or text into its own distinct block, use triple backticks
example:
  git status
  git commit -m

Featured Notebooks/Analysis/Deliverables

• Notebook/Markdown/Slide Deck Title
• Notebook/Markdown/Slide DeckTitle
• Blog Post
• Blog Post

Contributing Team Members

Team Lead (Contacts) : Nikhil Gupta

Other Members:

Blair Fernandes, [Asjad Baig]

Contact

• Feel free to contact me on [email protected] with any questions or if you are interested in contributing!