app.py

import streamlit as st
from sentence_transformers import SentenceTransformer, util
from transformers import (AutoModelForQuestionAnswering,
                          AutoTokenizer, pipeline)

import pandas as pd
import regex as re

# Select model for question answering
model_name = "deepset/roberta-base-squad2"

# Load model & tokenizer
model = AutoModelForQuestionAnswering.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Create pipeline
pipe = pipeline('question-answering', model=model_name, tokenizer=model_name)

# Load Harry Potter book corpus from link
book1_raw_0 = open("book_1.txt", mode="r", encoding="utf-8").read()

# Text pre-processing
# Remove page statements
book1_raw_1 = re.sub(r'Page \| [0-9]+ Harry Potter [a-zA-Z \-]+J.K. Rowling', '', book1_raw_0)

# Remove newlines
book1_raw_1 = re.sub(r'\n', '', book1_raw_1)

# Remove periods; this will relevant in the regrouping later
book1_raw_1 = re.sub(r'Mr. ', 'Mr ', book1_raw_1)
book1_raw_1 = re.sub(r'Ms. ', 'Ms ', book1_raw_1)
book1_raw_1 = re.sub(r'Mrs. ', 'Mrs ', book1_raw_1)

# Group into 6 sentences-long parts
paragraphs = re.findall("[^.?!]+[.?!][^.?!]+[.?!][^.?!]+[.?!][^.?!]+[.?!][^.?!]+[.?!][^.?!]+[.?!]", book1_raw_1)

st.title('Harry Potter and the Extractive Question Answering Model')

# Type in HP-related query here
query = st.text_area("Hello my dears! What is your question? Be patient please, I am not a Ravenclaw!")

if st.button('Accio Responsa!'):
    # Perform sentence embedding on query and sentence groups
    model_embed_name = 'sentence-transformers/msmarco-distilbert-dot-v5'
    
    model_embed = SentenceTransformer(model_embed_name)
    doc_emb = model_embed.encode(paragraphs)
    query_emb = model_embed.encode(query)
    
    #Compute dot score between query and all document embeddings
    scores = util.dot_score(query_emb, doc_emb)[0].cpu().tolist()
    
    #Combine docs & scores
    doc_score_pairs = list(zip(paragraphs, scores))
    
    #Sort by decreasing score and get only 3 most similar groups
    doc_score_pairs = sorted(doc_score_pairs, key=lambda x: x[1],
                         reverse=True)[:1]
    
    # Join these similar groups to form the context 
    context = "".join(x[0] for x in doc_score_pairs)

    # Perform the querying
    QA_input = {'question': query, 'context': context}
    res = pipe(QA_input)
    
    confidence = res.get('score')
    if confidence > 0.5:
        st.write(res.get('answer'))
    else:
        out = "Sorry dear, I'm not sure"
        st.write(out)
    #out = res.get('answer')