QnA-research-papers

About

A Question Answering Web application that takes in a question from a user and returns the top 6 answers according to the question. It uses a pretrained BERT QA model to perform extractive question answering on text.

Installation

1. clone the repo by running: git clone https://github.com/zeina99/QnA-research-papers.git

2. cd into the directory into the cloned repo

3. to install the requirements:

   1. make sure python is installed

   2. pip install -r requirements.txt

      OR

      pip install transformers flask torch scikit-learn numpy pdfplumber
      

4. run the app:

   python app.py

How it works:

Adding a pdf file:

PDF -> text extraction -> document ,paragraphs -> DB

Asking a question:

Question, pdf chosen -> top 6 paragraphs -> top 6 answers

1. Pdf to database

• when pdf gets added
  • text is extracted from the pdf
  • text is added to the database
  • the text is iterated over and gets divided into paragraphs. a paragraph is set to have a max of 250 words
  • list of paragraphs get added to database along with docuemnt-id they belong to

2. Ask a question

• When the desired pdf is selected and the question is entered, cosine-similarity is applied on all paragaraphs according to the pdf selected and the top 6 paragraphs that could hold the answer are returned.
• pretrained BERT QA model on 6 paragraphs
  • the pretrained BERT QA model is applied on all 6 paragraphs so we end up with 6 possible answers.

Technologies used:

• Pretrained question answering bert model
• sqlite db
• pdfplumber

Difficulties faced:

• pretrained BERT QA model doesnt work well on large amounts of text:
  • solution:
  • break down research papers into paragraphs and add them to the database
  • get the top 6 most similar paragraphs against the user query using cosine similarity
  • apply the pretrained BERT QA model on the top 6 paragraphs which results in 6 possible answers
• pdf plumber reads text from left to right, top to bottom and research papers have 2 columns of text
  • solution:
  • crop each page into two halves, extract text from left half then right half and combine them

Future improvements:

• find a way to ignore subscripts when extracting text from pdfs as it is causing problems
• divide paragraphs as present in pdf document, instead of a fixed word count
• implement a login system where each user can view their own pdfs

Screenshots

1. Adding a pdf:

2. Preparing a question to ask related to the pdf content:

• Question: What is the most competitive neural sequence transduction models have?
• Answer: an encoder-decoder structure

3. Getting top 6 answers according to the question.