parthplc / Extract-Information-using-Rule-Based-Methods

Information extraction Interview Task

Extract-Information-using-Rule-Based-Methods

Information extraction Interview Task

Folders

Score/Summary in csv folder
xml of pdf of research paper in xml folder
pdf of researchpaper in papaer folder

Task

1. Use TFIDF find most important topic(topics) and find all necessary sentences revolving round it using spacy matcher(find patters) --> ALgorithm.
2. Using Gorbid I converted PDF to XML which is saved in xml folder.
3. Using approach in task 1 found contributing sentences
4. Since even contributing sentences have length of about 8000 words used Bigbird Pegasus which can use more than 4000 tokens (unlike 576 in traditional transformer models)
5. Since Summeval have dependency issue for macos and kaggle/Colab I used the original rogue score library https://pypi.org/project/rouge-score/ to find metric of absractive summarization compared to original abstract.

All contributing sentences/ Rogue Score / Abstractive summarization is saved in csv folder file

Rogue Score

Final rogue score of bigbird model compared to abstract is

'rouge1': Score(precision=0.4372093023255814, recall=0.42152466367713004, fmeasure=0.4292237442922374)
'rouge2': Score(precision=0.14018691588785046, recall=0.13513513513513514, fmeasure=0.13761467889908258)
'rougeL': Score(precision=0.20465116279069767, recall=0.19730941704035873, fmeasure=0.2009132420091324)}

Code on Kaggle

Link : https://www.kaggle.com/code/parthplc/extract-data-from-research-paper-final?scriptVersionId=90756568 Score Csv Link : https://docs.google.com/spreadsheets/d/1l46_Zg5qS1aCgiTycvRk5EBTkq5tS4RRjnauUv4YOmc/edit?usp=sharing