Inspiration

Text summarizer can be used in various applications such as: quick overview of large set of text documents, search result summary for web pages, email classification/summarization... The classical example of how human summarizes text is best illustrated by the scientific papers: for a paper of 4-5 pages (around 5000 words), the owner often writes a abstract of that paper in short summary (around 250 words). Taking a large collection of scientific papers, we can develop a machine learning model that will be trained and eventually be able to generate good abstract from given text input (5000 words) the same way an author would write an abstract.

What it does

Given an input text (5000 words) the Text Summarizer will generated an abstract (250 words), similar to the way an abstract of a scientific paper is written.

How I built it

Firstly the text input (full paper text and abstract text) need to be converted into vectors (so called embedding). There are several pre-trainned Word2Vec model. This project considered using the Glove (Global Vectors for Word Representation). The dataset for training and testing comes from the CORE dataset. The CORE dataset contains millions of research output text (full text and abstract) and it can be freely download from the (https://core.ac.uk/services/dataset/) CORE project website. Then the training data is used for a conditional GAN which has a Generator and a Discriminator. The Generator takes a training full paper text and try to generate a corresponding abstract. The Discriminator will give feedback to the Generator on real abstract and generated abstract. Over the time, the Generator will learnt and improve the generated output, make the text similar to a real abstract.

Challenges I ran into

There are quite a number of tutorials, books about Tensorflow. However, much of the examples code is written with the previous version of Tensorflow. I am very new to Tensorflow, therefore it takes time to learn the tool and play around before I can start implement the Text Summarizer as I want it to be.

Accomplishments that I'm proud of

Learnt and apply several Machine Learning techniques with the new Tensorflow 2.0. Also Text Summarizer can be very useful in many applications.

What I learned

Try Colab notebook as soon as possible.It takes time to implement, adjust, or sometimes even change the model, hyper parameters, the ealier we starts with the code, the better because we will see the result quicker and adjust it accordingly.

What's next for Tensorflow Text Summarizer

The CORE dataset is very large (recently > 330 GB gz files and extracted size is around 1.3 TB). It would be interesting to fully train Text Summarizer on the full dataset and evaluate how the generated abstract improved.

Acknowledgement

This project would not be completed without the tutorial/data/work from sites listed below. My big thanks to the authors.

  1. Pix2Pix: Image to image translation using conditional GAN' https://www.tensorflow.org/alpha/tutorials/generative/pix2pix

  2. CORE Dataset: Millions of research output text (full text and abstract) https://core.ac.uk/services/dataset/

  3. GLOVE: Global Vectors for Word Representation https://nlp.stanford.edu/projects/glove/

  4. GloVe-as-a-TensorFlow-Embedding-Layer https://github.com/guillaume-chevalier/GloVe-as-a-TensorFlow-Embedding-Layer

Built With

Share this project:
×

Updates