PranshuKala / Text-Summarization

Text-Summarization Report

---

Data Collection

• Searched for suitable dataset to be trained on my hardware
  
• Selected 'Inshorts News Summary Dataset'
• Successfully loaded the selected dataset , saved that in Data Folder.

Data Preprocessing and EDA

• Extracted the required columns from the master dataset.
• Applied nlp preprocessing techniques on the extracted dataset.
• Visualized the data distribution between the input and target column.
• Saved the preprocessed dataset.

Model Building

Abstractive Approach

• Selected t5-small model for abstractive summarization.

• Prefine tuning results stored at Summarization_Model/model.ipynb .

• Fine tuned it on my dataset , evaluated results at Summarization_Model/evaluation.ipynb

• Model Size after fine tuning n ~ 800 MB .

• Final rogue scores after fine tuning are as follows: Rouge-1: 0.42 Rouge-2: 0.22 Rouge-L: 0.34

Extractive Approach

• Performed Extractive Summarization pre-processing on the dataset by removing stopwords, punctuation, and special characters and white spaces.

• Selected Text Rank Algorithm for extractive summarization .

• Implemented through py summa library.

• Rouge scores for extractive summarization are as follows: (for 20% ratio of input text vs summary length) Rouge-1: 0.25 Rouge-2: 0.12 Rouge-L: 0.19

• Optimized ratio Length of summary to input text to get better results.

Model Interface

• Built Model Interface using streamlit library.
• Implemented in Model_Interface/app.py.
• Visualized Abstract and Extractive Summarization results on the interface.