Deep Learning - Text Summarization
Goal of the project is to summarize the document as accurately as possible. This NLP task does not have a specific metric to describe accuracy of the model, hence metrics like Bleu Score and Rogue were used as an indicator.
Team: Pawanjeet Kaur, Shivam Duseja
Dataset: WikiHowAll.csv
Models: LSTM, LSTM with attention , GRU, GRU with attention, Stacked LSTM
Average time for one model to run on Colab Pro with corss validation : 4-5 hours
Best Bleu Score: 0.365
https://towardsdatascience.com/text-summarization-using-deep-neural-networks-e7ee7521d804#a87e-1aa55d239c9e
Why we chose Document Summarizer?
Increasing availability of online information has added more importance on the intensive research in the area of automatic text summarization within the Natural Language Processing (NLP) community. Text Summarization is basically an algorithm that shortens long pieces of sentences to produce a coherent summary that contains only the major points or the gist of the sentence which in turn saves time and helps us to identify the interest area. In this work, we build an abstractive text summarizer for different news articles using deep neural networks (Encoder-Decoder architecture).
For nearly the last half century, the field of summarization has been investigated by the NLP community. Earlier research works, proposed extraction of salient sentences from text using features like - phrase/word frequency (Luhn, 1958), position in the text (Baxendale, 1958) and key phrases (Edmundson, 1969).
Depending on the type of summary to be generated, different approaches of summarization have been proposed. The two major approaches for text summarization are - extractive and the abstractive technique. The extractive technique is mainly concerned with the extraction of words/phrases from the input sentence. On the other hand, the abstractive summary technique puts strong emphasis on the form - aiming to generate a grammatical summary thereby requiring advanced language modelling techniques.
Practical work after choosing the problem and dataset:
1. Data Exploration and Pre processing ( NLP task)
2. Exploratory Data Analysis - on sentence length and checking distribution of the data
3. Deep Model Design :
Seq2seq Model with Attention using LSTM ,
Seq2seq Model with Attention using GRU and Teacher Forcing
Three Stacked LSTM
4. Metric Calculation - Rouge , Bleu and F-Score
All these steps are described in detail in report attached. My teammate recently published his article on this project on toward data science.
Public Link: https://towardsdatascience.com/text-summarization-using-deep-neural-networks-e7ee7521d804#a87e-1aa55d239c9e