app.py

import time
import pandas as pd

import torch
from keybert import KeyBERT
from sentence_transformers import SentenceTransformer
from keyphrase_vectorizers import KeyphraseCountVectorizer
from transformers import T5ForConditionalGeneration,T5Tokenizer

import nltk
from nltk.tokenize import sent_tokenize

from huggingface_hub import snapshot_download, HfFolder
import streamlit as st

import traceback
import logging


nltk.download('stopwords')
nltk.download('punkt')

logger = logging.getLogger(__name__)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

HfFolder.save_token(st.secrets["hf-auth-token"])


@st.cache(allow_output_mutation=True)
def load_model():
    try:
        # Load KeyBert Model
        tmp_model = SentenceTransformer('valurank/MiniLM-L6-Keyword-Extraction', use_auth_token=True)
        kw_extractor = KeyBERT(tmp_model)
        
        # Load T5 for Paraphrasing
        t5_model = T5ForConditionalGeneration.from_pretrained('valurank/t5-paraphraser', use_auth_token=True)
        t5_tokenizer = T5Tokenizer.from_pretrained('t5-base')
        t5_model = t5_model.to(device)
        return kw_extractor, t5_model, t5_tokenizer
    except Exception:
        st.error('Error Loading Models. Please contact admin')
        logger.error(traceback.format_exc())
    
    
kw_extractor, t5_model, t5_tokenizer = load_model()


@st.cache()
def get_keybert_results_with_vectorizer(text, number_of_results=20):

    try:
        keywords = kw_extractor.extract_keywords(text, vectorizer=KeyphraseCountVectorizer(), stop_words=None, top_n=number_of_results)
        return keywords
    except Exception:
        st.error('Error running Keybert. Please contact admin')
        logger.error(traceback.format_exc())



@st.cache()
def t5_paraphraser(text, number_of_results=5):
    try:
        text =  "paraphrase: " + text + " </s>"
        max_len = 2048
        encoding = t5_tokenizer.encode_plus(text, pad_to_max_length=True, return_tensors="pt")
        input_ids, attention_masks = encoding["input_ids"].to(device), encoding["attention_mask"].to(device)
    
        beam_outputs = t5_model.generate(
            input_ids=input_ids, attention_mask=attention_masks,
            do_sample=True,
            max_length=2048,
            top_k=50,
            top_p=0.95,
            early_stopping=True,
            num_return_sequences=number_of_results
        )
        
        final_outputs =[]
        for beam_output in beam_outputs:
            sent = t5_tokenizer.decode(beam_output, skip_special_tokens=True, clean_up_tokenization_spaces=True)
            final_outputs.append(sent)
        return final_outputs
    except Exception:
        st.error('Error running T5 Paraphrasing. Please contact admin')
        logger.error(traceback.format_exc())
    
    
    
#### Extract Sentences with Keywords -> Paraphrase multiple versions -> Extract Keywords again
def extract_paraphrased_sentences(article):
    try: 
        start1 = time.time()
        with st.spinner('Extraction Keywords from Original Document...'):
            original_keywords = get_keybert_results_with_vectorizer(article)
        
            article_sentences = sent_tokenize(article)
            target_sentences = [sent for sent in article_sentences if any(kw[0] in sent for kw in original_keywords)]        
        st.success('Keyword Extraction from Original Document finished in {}'.format(time.time() - start1))
   
        start2 = time.time()
        with st.spinner('Extracting Keywords from Paraphrased Target Sentences...'):
            t5_paraphrasing_keywords = []
        
            for sent in target_sentences:
                ### T5
                t5_paraphrased = t5_paraphraser(sent)
                t5_keywords = [get_keybert_results_with_vectorizer(i) for i in t5_paraphrased]
                t5_keywords = [(word[0], word[1]) for s in t5_keywords for word in s]
        
                t5_paraphrasing_keywords.extend(t5_keywords)
        st.success('Keyword Extraction from Paraphrased Target Sentences finished in {}'.format(time.time() - start2))
    
        original_keywords_df = pd.DataFrame(original_keywords, columns=['Keyword', 'Score'])
        
        t5_keywords_df = pd.DataFrame(t5_paraphrasing_keywords, columns=['Keyword', 'Score']).sort_values(by='Score', ascending=False).drop_duplicates(subset=['Keyword'], keep='first').reset_index(drop=True)
    
        unique_keywords_df = pd.DataFrame([i for i in t5_paraphrasing_keywords if not original_keywords_df['Keyword'].str.contains(i[0]).any()], columns=['Keyword', 'Score']).sort_values(by='Score', ascending=False).drop_duplicates(subset=['Keyword'], keep='first').reset_index(drop=True)
    
        total_end = time.time()-start1
    
        return t5_keywords_df, original_keywords_df, unique_keywords_df, total_end
    except Exception:
        st.error('Error running Extraction Pipeline. Please contact admin')
        logger.error(traceback.format_exc())    

doc = st.text_area("Enter a custom document")

if doc:
    t5_keywords_df, original_keywords_df, unique_keywords_df, total_end = extract_paraphrased_sentences(doc)
    
    # extract_paraphrased_article(input_list[0])
    st.text(f'T5 PARAPHRASING RUNTIME: {total_end}\n')
    
    st.subheader('\nOriginal Keywords Extracted:\n\n')
    st.dataframe(original_keywords_df)
    
    st.subheader('\nT5 Unique New Keywords Extracted:\n\n')
    st.dataframe(unique_keywords_df)
       
    st.subheader('\nT5 Keywords Extracted:\n\n')
    st.dataframe(t5_keywords_df)