app.py

#loading tfidf dataset 
import pandas as pd 
newsdf_sample = pd.read_excel("complete_tfidf_25.xlsx",engine="openpyxl")

print("file size",len(newsdf_sample))

#preprocessing for better tokenization (needed for tfidf)
import nltk
nltk.download('stopwords')
nltk.download('wordnet')
nltk.download('omw-1.4')
from nltk.corpus import stopwords
stopwords_list = stopwords.words('english')
stopwords_list

def process_row(row):
    import re
    from textblob import Word
    from string import punctuation
    from nltk.stem.snowball import SnowballStemmer    

    
    #Mail address
    row = re.sub('(\S+@\S+)(com|\s+com)', ' ', row)
    
    #Username
    row = re.sub('(\S+@\S+)', ' ', row)
#     print('username',len(row.split()))
    
    #punctuation
    punctuation = punctuation + '\n' + '—“,”‘-’' + '0123456789' +"\t"
    row = ''.join(word for word in row if word not in punctuation)
#     print('punctuation',len(row.split()))
#     print('punctuation',row)
    
    #Lower case 
    row = row.lower()
#     print('lower',len(row.split()))
    
    #Stopwords
    stop = stopwords_list
    row = ' '.join(word for word in row.split() if word not in stop ) 
#     print('stop',len(row.split()))
#     print('stop',row)
    
#     Lemma
    row = " ".join([Word(word).lemmatize() for word in row.split()])
#     print('lemma',len(row.split()))
#     print('lemma',row)
    
    #Stemming
    stemmer = SnowballStemmer(language='english')
    row = " ".join([stemmer.stem(word) for word in row.split()])
#     print('stem',len(row.split()))
#     print('stem',row)
    
    #Extra whitespace
    row = re.sub('\s{1,}', ' ', row)
#     print('extra white',len(row.split()))
    
    row = " ".join([word for word in row.split() if len(word) > 2])

    return row

import pickle
kmeans_tfidf = pickle.load( open( "kmeans_tfidf_25_complete.p", "rb" ) )
vectorizer = pickle.load(open("tfidf_vectorizer_complete.p","rb"))

import matplotlib.pyplot as plt
from wordcloud import WordCloud


dictt_cluster_words={}

for i in range(0,20):
  # print(i)
  temp_df = newsdf_sample[newsdf_sample.exp25==i]
  text_list= temp_df["tfidf_cleaned"].values
  text_list = [element for element in text_list if str(element) != "nan"]
  single_text = " ".join(text_list)
  wordcloud = WordCloud(width = 1000, height = 500, max_words=1000).generate(single_text)
  dictt_cluster_words[i] = wordcloud.words_


#summarization model


from transformers import PegasusForConditionalGeneration, PegasusTokenizer
from transformers import pipeline
import torch



model_name = 'google/pegasus-cnn_dailymail'
device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = PegasusTokenizer.from_pretrained(model_name)
model = PegasusForConditionalGeneration.from_pretrained(model_name).to(device)

def return_summary(text):
  src_text =[text]
  batch = tokenizer(src_text, truncation=True, padding="longest", return_tensors="pt").to(device)
  translated = model.generate(**batch)
  tgt_text = tokenizer.batch_decode(translated, skip_special_tokens=True)
  tgt_text= tgt_text[0].replace("<n>"," ")
  return tgt_text

############



def return_squad_answer(question, relevant_text):

  qa_pipeline = pipeline(
    "question-answering",
    model="mvonwyl/distilbert-base-uncased-finetuned-squad2",#csarron/bert-base-uncased-squad-v1",
    tokenizer="mvonwyl/distilbert-base-uncased-finetuned-squad2",#csarron/bert-base-uncased-squad-v1"
  )

  predictions = qa_pipeline({
    'context': relevant_text,
    'question': question
  })

  print(predictions)
  return predictions["answer"]

#keyword based cluster selection would be better
#document selection based on tfidf vector

import numpy as np
import math
def l2_norm(a):
    return math.sqrt(np.dot(a,a))

def cosine_similarity(a,b):
    return abs(np.dot(a,b)/ (l2_norm(a) * l2_norm(b)))

def return_selected_cluster(ques):
  ques_clean = process_row(ques)
  cluster_selected =-1
  cluster_score =0
  for clus_id in dictt_cluster_words:
    score_temp=0
    for word in ques_clean.split():
      dictt_temp = dictt_cluster_words[clus_id]    
      if word in dictt_temp:
        score_temp+=dictt_temp[word]
    if score_temp>cluster_score:
      cluster_selected = clus_id
      cluster_score = score_temp
  return cluster_selected


def get_summary_answer(Question):
  print("question: ", Question)
  cluster_selected = return_selected_cluster(Question)

  temp_df = newsdf_sample[newsdf_sample.exp25==cluster_selected]
  tfidf_ques = vectorizer.transform([process_row(Question)]).todense()
  cosine_score = []
  for sent in temp_df["tfidf_cleaned"].values:
    val = vectorizer.transform([sent]).todense()
    # print(np.array(tfidf_ques)[0], np.array(val)[0])
    cos_score = cosine_similarity(np.array(tfidf_ques)[0],np.array(val)[0])
    cosine_score.append(cos_score)

  temp_df["cos_score"] = cosine_score
  temp_df = temp_df.sort_values(by=['cos_score'], ascending=False)

  relevant_docs = temp_df["cleaned_doc"][:20]
  relevant_text = " ".join(relevant_docs)
  print("relevant_text", relevant_text)

  # print("summary - ",return_summary(relevant_text))
  # print("squad answer- ",return_squad_answer(ques, relevant_text))

  summary = return_summary(relevant_text)
  squad_answer = return_squad_answer(Question, relevant_text)

  relevant_text = " ".join(relevant_text.split()[:min(250,len(relevant_text.split()))])


  return relevant_text, summary, squad_answer


import gradio as gr
iface = gr.Interface(fn = get_summary_answer,
                     inputs = gr.Textbox(type="text", label="Type your question"), 
                    #  outputs = ["text", "text"], 
                     outputs = [
                         gr.Textbox(type="text", value=1, label="Relevant text"),
                         gr.Textbox(type="text", value=2, label="Answer from Generative Model"),
                         gr.Textbox(type="text", value=3, label="Answer from SQuAD model"),      
                     ],
                     title = "20NewsGroup_QA", 
                     description ="Returns answer from 20NewsGroup dataset")
iface.launch(inline = False, debug = True)