app.py

import streamlit as st
from transformers import pipeline
import re
import time
import requests


HF_SPACES_API_KEY = st.secrets["HF_token"]

API_URL = "https://api-inference.huggingface.co/models/microsoft/prophetnet-large-uncased-squad-qg"
headers = {"Authorization": HF_SPACES_API_KEY}

def query(payload):
	response = requests.post(API_URL, headers=headers, json=payload)
	return response.json()


#-----------------------------------------------------------

API_URL_evidence ="https://api-inference.huggingface.co/models/google/flan-t5-xxl"
headers_evidence = {"Authorization": HF_SPACES_API_KEY}

def query_evidence(payload):
	response = requests.post(API_URL_evidence, headers=headers_evidence, json=payload)
	return response.json()

#-----------------------------------------------------------
st.title('Welcome to :blue[FACTIFY - 5WQA] ') 
st.header('5W Aspect-based Fact Verification through Question Answering :blue[Web Demo]')
st.subheader('Here are a few steps to begin exploring and interacting with this demo.')
st.caption('First you need to input your claim.')
st.caption('Then you need to input your evidence and press **:green[Submit]**.')
st.caption('Upon completing these two steps, kindly wait for a minute to receive the results.')

st.caption('Start by inputting the following instance of a claim and corresponding evidence into the designated text fields.')
#-----------------------------------------------------------------------------------------------
st.caption('**Example 1**')
st.caption(''':green[Claim:] :point_right: Amazon announced on March 16 it would hire 100,000 new warehouse and delivery workers and 
raise wages $ 2 per hour through April in response to increased demand for its services because of the coronavirus pandemic .''')

st.caption(''':green[Evidence:] :point_right: Due to the consumers increasingly relying on online retailers, 
Amazon planned to hire over 99,000 workers in the warehouse and delivery sector during the Pandemic in the USA.''')

#-----------------------------------------------------------------------------------------------
st.caption('**Example 2**')
st.caption(''':green[Claim:] :point_right: In China, Buddhist monks and nuns lived together in places such as the Yunnan monastery.''')

st.caption(''':green[Evidence:] :point_right: Monastics in Japan are particularly exceptional in the Buddhist tradition because the monks and nuns can marry after receiving their higher ordination . 	''')

#-----------------------------------------------------------------------------------------------
st.caption('**Example 3**')
st.caption(''':green[Claim:] :point_right: In Batman, Penguin hydrates the henchmen with water contaminated with atomic waste.''')

st.caption(''':green[Evidence:] :point_right: And Penguin even schemes his way into the Batcave along with five dehydrated henchmen ; 
this plan fails when the henchmen are unexpectedly killed 
when he mistakenly rehydrates them with heavy water contaminated with atomic waste , 
regularly used to recharge the Batcave s atomic pile . 	''')

#-----------------------------------------------------------
def proc():
    st.write(st.session_state.text_key)

# st.text_area('enter text', on_change=proc, key='text_key')


claim_text=st.text_area("Enter your claim:", on_change=proc, key='text_key')

# form_claim = st.form(key='my_claim')
# form_claim.text_input(label='Enter your claim')
# claim_text = form_claim.form_submit_button(label='Submit')




# evidence_text=st.text_area("Enter your evidence:")

form_evidence = st.form(key='my_evidence')
form_evidence.text_input(label='Enter your evidence')
evidence_text = form_evidence.form_submit_button(label='Submit')

if evidence_text:
    st.caption(':green[Kindly hold on for a few minutes while the QA pairs are being generated]')
    st.caption(':blue[At times, you may encounter null/none outputs, which could be a result of a delay in loading the models through the API. If you experience this problem, kindly try again after a few minutes.]')


import pandas as pd
from rouge_score import rouge_scorer
import numpy as np
from allennlp.predictors.predictor import Predictor
import allennlp_models.tagging
predictor = Predictor.from_path("structured-prediction-srl-bert.tar.gz")

#---------------------------------------------------------------
def claim(text):
    import re
    def remove_special_chars(text):
        # Remove special characters that are not in between numbers
        text = re.sub(r'(?<!\d)[^\w\s]+(?!\d)', '', text)

        return text

    df = pd.DataFrame({'claim' : remove_special_chars(text)},index=[0])

    def srl_allennlp(sent):
      try:
        #result = predictor.predict(sentence=sent)['verbs'][0]['description']
        #result = predictor.predict(sentence=sent)['verbs'][0]['tags']
        result = predictor.predict(sentence=sent)
        return(result)
      except IndexError: 
        pass
      #return(predictor.predict(sentence=sent))

    df['allennlp_srl'] = df['claim'].apply(lambda x: srl_allennlp(x))

    df['number_of_verbs'] = ''
    df['verbs_group'] = ''
    df['words'] = ''
    df['verbs'] = ''
    df['modified'] =''

    col1 = df['allennlp_srl']
    for i in range(len(col1)):
      num_verb = len(col1[i]['verbs'])
      df['number_of_verbs'][i] = num_verb
      df['verbs_group'][i] = col1[i]['verbs']
      df['words'][i] = col1[i]['words']

      x=[]
      for verb in range(len(col1[i]['verbs'])):
        x.append(col1[i]['verbs'][verb]['verb'])
      df['verbs'][i] = x

      verb_dict ={}
      desc = []
      for j in range(len(col1[i]['verbs'])):
        string = (col1[i]['verbs'][j]['description'])
        string = string.replace("ARG0", "who")
        string = string.replace("ARG1", "what")
        string = string.replace("ARGM-TMP", "when")
        string = string.replace("ARGM-LOC", "where")
        string = string.replace("ARGM-CAU", "why")
        desc.append(string)
        verb_dict[col1[i]['verbs'][j]['verb']]=string
      df['modified'][i] = verb_dict


    #----------FOR COLUMN "WHO"------------#
    df['who'] = ''
    for j in range(len(df['modified'])):
      val_list = []
      val_string = ''
      for k,v in df['modified'][j].items():
        # print(type(v))
        val_list.append(v)

      who = []
      for indx in range(len(val_list)):
        val_string = val_list[indx]
        pos = val_string.find("who: ")
        substr = ''

        if pos != -1:
          for i in range(pos+5, len(val_string)):
            if val_string[i] == "]":
              break
            else:
              substr = substr + val_string[i]
        else:
          pass
        if len(substr)!= 0: 
            who.append(substr)
        else:
            pass
        
 
    df['who'][j] = "<sep>".join(who)
#     else:
#         continue
    #----------FOR COLUMN "WHAT"------------#
    df['what'] = ''
    for j in range(len(df['modified'])):
      val_list = []
      val_string = ''
      for k,v in df['modified'][j].items():
        # print(type(v))
        val_list.append(v)

      what = []
      for indx in range(len(val_list)):
        val_string = val_list[indx]
        pos = val_string.find("what: ")
        substr = ''

        if pos != -1:
          for i in range(pos+6, len(val_string)):
            if val_string[i] == "]":
              break
            else:
              substr = substr + val_string[i]
        else:
          pass
        if len(substr)!= 0: 
            what.append(substr)
        else:
            pass
  
    df['what'][j] = "<sep>".join(what)
#     else:
#         continue

    #----------FOR COLUMN "WHY"------------#
    df['why'] = ''
    for j in range(len(df['modified'])):
      val_list = []
      val_string = ''
      for k,v in df['modified'][j].items():
        # print(type(v))
        val_list.append(v)

      why = []
      for indx in range(len(val_list)):
        val_string = val_list[indx]
        pos = val_string.find("why: ")
        substr = ''

        if pos != -1:
          for i in range(pos+5, len(val_string)):
            if val_string[i] == "]":
              break
            else:
              substr = substr + val_string[i]
        else:
          pass
        if len(substr)!= 0: 
            why.append(substr)
        else:
            pass

    df['why'][j] = "<sep>".join(why)
#     else:
#         continue

     #----------FOR COLUMN "WHEN"------------#
    df['when'] = ''
    for j in range(len(df['modified'])):
      val_list = []
      val_string = ''
      for k,v in df['modified'][j].items():
        # print(type(v))
        val_list.append(v)

      when = []
      for indx in range(len(val_list)):
        val_string = val_list[indx]
        pos = val_string.find("when: ")
        substr = ''

        if pos != -1:
          for i in range(pos+6, len(val_string)):
            if val_string[i] == "]":
              break
            else:
              substr = substr + val_string[i]
        else:
          pass
        if len(substr)!= 0: 
            when.append(substr)
        else:
            pass

    df['when'][j] = "<sep>".join(when)
#     else:
#         continue


    #----------FOR COLUMN "WHERE"------------#
    df['where'] = ''
    for j in range(len(df['modified'])):
      val_list = []
      val_string = ''
      for k,v in df['modified'][j].items():
        # print(type(v))
        val_list.append(v)

      where = []
      for indx in range(len(val_list)):
        val_string = val_list[indx]
        pos = val_string.find("where: ")
        substr = ''

        if pos != -1:
          for i in range(pos+7, len(val_string)):
            if val_string[i] == "]":
              break
            else:
              substr = substr + val_string[i]
        else:
          pass
        if len(substr)!= 0: 
            where.append(substr)
        else:
            pass

    df['where'][j] = "<sep>".join(where)


    data=df[["claim","who","what","why","when","where"]].copy()    
    return data
#-------------------------------------------------------------------------
def split_ws(input_list, delimiter="<sep>"):
    output_list = []
    for item in input_list:
        split_item = item.split(delimiter)
        for sub_item in split_item:
            sub_item = sub_item.strip()
            if sub_item:
                output_list.append(sub_item)
    return output_list

#--------------------------------------------------------------------------    
def calc_rouge_l_score(list_of_evidence, list_of_ans):
    scorer = rouge_scorer.RougeScorer(['rougeL'], use_stemmer=True)
    scores = scorer.score(' '.join(list_of_evidence), ' '.join(list_of_ans))
    return scores['rougeL'].fmeasure
#-------------------------------------------------------------------------
    

def rephrase_question_who(question):
    if not question.lower().startswith("who"):
        words = question.split()
        words[0] = "Who"
        return " ".join(words)
    else:
        return question
#------------------------------------------------------------------------    
def gen_qa_who(df):
    list_of_ques_who=[]
    list_of_ans_who=[]
    list_of_evidence_answer_who=[]
    rouge_l_scores=[]
    for i,row in df.iterrows():
        srl=df["who"][i]
        claim=df['claim'][i]
        answer= split_ws(df["who"])
        evidence=df["evidence"][i]
        #time.sleep(10)
        if srl!="":
            try:
                for j in range(0,len(answer)):
                    FACT_TO_GENERATE_QUESTION_FROM = f"""{answer[j]} [SEP] {claim}"""
                    #FACT_TO_GENERATE_QUESTION_FROM = f"""generate_who_based_question_from_context_using_the_next_answer:{answer[j]} [SEP] context:{claim}"""
                    #time.sleep(10)
                    question_ids = query({"inputs":FACT_TO_GENERATE_QUESTION_FROM, 
                              "num_beams":5, 
                              "early_stopping":True,
                                         "min_length": 100,"wait_for_model":True})[0]['generated_text'].capitalize()
                    question_ids = rephrase_question_who(question_ids)
                    list_of_ques_who.append(f"""Q{j+1}:{question_ids}""")
                    list_of_ans_who.append(f"""Ans{j+1}:{answer[j]}""")
                    input_evidence = f"answer_the_next_question_from_context: {question_ids} context: {evidence}"
                    #time.sleep(10)
                    answer_evidence = query_evidence({"inputs":input_evidence,"truncation":True,"wait_for_model":True})[0]['generated_text']
                    if answer_evidence.lower() in evidence.lower():
                        list_of_evidence_answer_who.append(f"""Evidence{j+1}:{answer_evidence}""")
                    else:
                        answer_evidence=""
                        list_of_evidence_answer_who.append(f"""No mention of 'who'in any related documents.""")
                    threshold = 0.2
                    list_of_pairs = [(answer_evidence, answer[j])]
                    rouge_l_score = calc_rouge_l_score(answer_evidence, answer[j])
                    if rouge_l_score >= threshold:
                        verification_status = 'Verified Valid'
                    elif rouge_l_score == 0:
                        verification_status = 'Not verifiable'
                    else:
                        verification_status = 'Verified False'
                    rouge_l_scores.append(verification_status)
            except:
                pass
        else:
            list_of_ques_who="No claims"
            list_of_ans_who=""
            list_of_evidence_answer_who="No mention of 'who'in any related documents."
            rouge_l_scores="Not verifiable"
    return list_of_ques_who,list_of_ans_who,rouge_l_scores,list_of_evidence_answer_who
#------------------------------------------------------------    

def rephrase_question_what(question):
    if not question.lower().startswith("what"):
        words = question.split()
        words[0] = "What"
        return " ".join(words)
    else:
        return question
#----------------------------------------------------------        
def gen_qa_what(df):
    list_of_ques_what=[]
    list_of_ans_what=[]
    list_of_evidence_answer_what=[]
    rouge_l_scores=[]
    for i,row in df.iterrows():
        srl=df["what"][i]
        claim=df['claim'][i]
        answer= split_ws(df["what"])
        evidence=df["evidence"][i]
        #time.sleep(10)
        if srl!="":
            try:
                for j in range(0,len(answer)):
                    FACT_TO_GENERATE_QUESTION_FROM = f"""{answer[j]} [SEP] context:{claim}"""
                    #time.sleep(10)
                    question_ids = query({"inputs":FACT_TO_GENERATE_QUESTION_FROM, 
                              "num_beams":5, 
                              "early_stopping":True,
                                         "min_length": 100,"wait_for_model":True})[0]['generated_text'].capitalize()
                    question_ids = rephrase_question_what(question_ids)
                    list_of_ques_what.append(f"""Q{j+1}:{question_ids}""")
                    list_of_ans_what.append(f"""Ans{j+1}:{answer[j]}""")
                    input_evidence = f"answer_the_next_question_from_context: {question_ids} context: {evidence}"
                    #time.sleep(10)
                    answer_evidence = query_evidence({"inputs":input_evidence,"truncation":True,"wait_for_model":True})[0]['generated_text']
                    if answer_evidence.lower() in evidence.lower():
                        list_of_evidence_answer_what.append(f"""Evidence{j+1}:{answer_evidence}""")
                        
                    else:
                        answer_evidence=""
                        list_of_evidence_answer_what.append(f"""No mention of 'what'in any related documents.""")
                    threshold = 0.2
                    list_of_pairs = [(answer_evidence, answer[j])]
                    rouge_l_score = calc_rouge_l_score(answer_evidence, answer[j])
                    if rouge_l_score >= threshold:
                        verification_status = 'Verified Valid'
                    elif rouge_l_score == 0:
                        verification_status = 'Not verifiable'
                    else:
                        verification_status = 'Verified False'
                    rouge_l_scores.append(verification_status)
            except:
                pass
        else:
            list_of_ques_what="No claims"
            list_of_ans_what=""
            list_of_evidence_answer_what="No mention of 'what'in any related documents."
            rouge_l_scores="Not verifiable"
    return list_of_ques_what,list_of_ans_what,rouge_l_scores,list_of_evidence_answer_what
#----------------------------------------------------------    

def rephrase_question_why(question):
    if not question.lower().startswith("why"):
        words = question.split()
        words[0] = "Why"
        return " ".join(words)
    else:
        return question

#---------------------------------------------------------        
def gen_qa_why(df):
    list_of_ques_why=[]
    list_of_ans_why=[]
    list_of_evidence_answer_why=[]
    rouge_l_scores=[]
    for i,row in df.iterrows():
        srl=df["why"][i]
        claim=df['claim'][i]
        answer= split_ws(df["why"])
        evidence=df["evidence"][i]
        #time.sleep(10)
        if srl!="":
            try:
                for j in range(0,len(answer)):
                    FACT_TO_GENERATE_QUESTION_FROM = f"""{answer[j]} [SEP] {claim}"""
                    #time.sleep(10)
                    question_ids = query({"inputs":FACT_TO_GENERATE_QUESTION_FROM, 
                              "num_beams":5, 
                              "early_stopping":True,
                                         "min_length": 100,"wait_for_model":True})[0]['generated_text'].capitalize()
                    question_ids = rephrase_question_why(question_ids)
                    list_of_ques_why.append(f"""Q{j+1}:{question_ids}""")
                    list_of_ans_why.append(f"""Ans{j+1}:{answer[j]}""")
                    input_evidence = f"answer_the_next_question_from_context: {question_ids} context: {evidence}"
                    #time.sleep(10)
                    answer_evidence = query_evidence({"inputs":input_evidence,"truncation":True,"wait_for_model":True})[0]['generated_text']
                    if answer_evidence.lower() in evidence.lower():
                        list_of_evidence_answer_why.append(f"""Evidence{j+1}:{answer_evidence}""")
                    else:
                        answer_evidence=""
                        list_of_evidence_answer_why.append(f"""No mention of 'why'in any related documents.""")
                    threshold = 0.2
                    list_of_pairs = [(answer_evidence, answer[j])]
                    rouge_l_score = calc_rouge_l_score(answer_evidence, answer[j])
                    if rouge_l_score >= threshold:
                        verification_status = 'Verified Valid'
                    elif rouge_l_score == 0:
                        verification_status = 'Not verifiable'
                    else:
                        verification_status = 'Verified False'
                    rouge_l_scores.append(verification_status)
            except:
                pass
        else:
            list_of_ques_why="No claims"
            list_of_ans_why=""
            list_of_evidence_answer_why="No mention of 'why'in any related documents."
            rouge_l_scores="Not verifiable"
    return list_of_ques_why,list_of_ans_why,rouge_l_scores,list_of_evidence_answer_why

#---------------------------------------------------------    

def rephrase_question_when(question):
    if not question.lower().startswith("when"):
        words = question.split()
        words[0] = "When"
        return " ".join(words)
    else:
        return question
#---------------------------------------------------------        
def gen_qa_when(df):
    list_of_ques_when=[]
    list_of_ans_when=[]
    list_of_evidence_answer_when=[]
    rouge_l_scores=[]
    for i,row in df.iterrows():
        srl=df["when"][i]
        claim=df['claim'][i]
        answer= split_ws(df["when"])
        evidence=df["evidence"][i]
        #time.sleep(10)
        if srl!="":
            try:
                for j in range(0,len(answer)):
                    FACT_TO_GENERATE_QUESTION_FROM = f"""{answer[j]} [SEP] {claim}"""
                    #time.sleep(10)
                    question_ids = query({"inputs":FACT_TO_GENERATE_QUESTION_FROM, 
                              "num_beams":5, 
                              "early_stopping":True,
                                         "min_length": 100,"wait_for_model":True})[0]['generated_text'].capitalize()
                    question_ids = rephrase_question_when(question_ids)
                    list_of_ques_when.append(f"""Q{j+1}:{question_ids}""")
                    list_of_ans_when.append(f"""Ans{j+1}:{answer[j]}""")
                    input_evidence = f"answer_the_next_question_from_context: {question_ids} context: {evidence}"
                    #time.sleep(10)
                    answer_evidence = query_evidence({"inputs":input_evidence,"truncation":True,"wait_for_model":True})[0]['generated_text']
                    if answer_evidence.lower() in evidence.lower():
                        list_of_evidence_answer_when.append(f"""Evidence{j+1}:{answer_evidence}""")
                    else:
                        answer_evidence=""
                        list_of_evidence_answer_when.append(f"""No mention of 'when'in any related documents.""")
                    threshold = 0.2
                    list_of_pairs = [(answer_evidence, answer[j])]
                    rouge_l_score = calc_rouge_l_score(answer_evidence, answer[j])
                    if rouge_l_score >= threshold:
                        verification_status = 'Verified Valid'
                    elif rouge_l_score == 0:
                        verification_status = 'Not verifiable'
                    else:
                        verification_status = 'Verified False'
                    rouge_l_scores.append(verification_status)
            except:
                pass
        else:
            list_of_ques_when="No claims"
            list_of_ans_when=""
            list_of_evidence_answer_when="No mention of 'when'in any related documents."
            rouge_l_scores="Not verifiable"
    return list_of_ques_when,list_of_ans_when,rouge_l_scores,list_of_evidence_answer_when

#------------------------------------------------------   

def rephrase_question_where(question):
    if not question.lower().startswith("where"):
        words = question.split()
        words[0] = "Where"
        return " ".join(words)
    else:
        return question
#------------------------------------------------------      
def gen_qa_where(df):
    list_of_ques_where=[]
    list_of_ans_where=[]
    list_of_evidence_answer_where=[]
    rouge_l_scores=[]
    for i,row in df.iterrows():
        srl=df["where"][i]
        claim=df['claim'][i]
        answer= split_ws(df["where"])
        evidence=df["evidence"][i]
        #time.sleep(10)
        if srl!="":
            try:
                for j in range(0,len(answer)):
                    FACT_TO_GENERATE_QUESTION_FROM = f"""{answer[j]} [SEP] {claim}"""
                    #time.sleep(10)
                    question_ids = query({"inputs":FACT_TO_GENERATE_QUESTION_FROM, 
                              "num_beams":5, 
                              "early_stopping":True,
                                         "min_length": 100,"wait_for_model":True})[0]['generated_text'].capitalize()
                    question_ids = rephrase_question_where(question_ids)
                    list_of_ques_where.append(f"""Q{j+1}:{question_ids}""")
                    list_of_ans_where.append(f"""Ans{j+1}:{answer[j]}""")
                    input_evidence = f"answer_the_next_question_from_context: {question_ids} context: {evidence}"
                    #time.sleep(10)
                    answer_evidence = query_evidence({"inputs":input_evidence,"truncation":True,"wait_for_model":True})[0]['generated_text']
                    if answer_evidence.lower() in evidence.lower():
                        list_of_evidence_answer_where.append(f"""Evidence{j+1}:{answer_evidence}""")
                    else:
                        answer_evidence=""
                        list_of_evidence_answer_where.append(f"""No mention of 'where'in any related documents.""")
                    threshold = 0.2
                    list_of_pairs = [(answer_evidence, answer[j])]
                    rouge_l_score = calc_rouge_l_score(answer_evidence, answer[j])
                    if rouge_l_score >= threshold:
                        verification_status = 'Verified Valid'
                    elif rouge_l_score == 0:
                        verification_status = 'Not verifiable'
                    else:
                        verification_status = 'Verified False'
                    rouge_l_scores.append(verification_status)
            except:
                pass
        else:
            list_of_ques_where="No claims"
            list_of_ans_where=""
            list_of_evidence_answer_where="No mention of 'where'in any related documents."
            rouge_l_scores="Not verifiable"
    return list_of_ques_where,list_of_ans_where,rouge_l_scores,list_of_evidence_answer_where

#------------------------------------------------------    


#------------------------------------------------------------  
if claim_text:
    if evidence_text:
        df=claim(claim_text)
        df["evidence"]=evidence_text
        final_df = pd.DataFrame(columns=['Who Claims', 'What Claims', 'When Claims', 'Where Claims', 'Why Claims'])
        final_df["Who Claims"]=gen_qa_who(df)
        final_df["What Claims"]=gen_qa_what(df)
        final_df["When Claims"]=gen_qa_when(df)
        final_df["Where Claims"]=gen_qa_where(df)
        final_df["Why Claims"]=gen_qa_why(df)
        st.dataframe(final_df)