app.py

import joblib
import streamlit as st
import json
import requests
from bs4 import BeautifulSoup
from datetime import date
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing.text import Tokenizer
from tensorflow.keras.preprocessing.sequence import pad_sequences
import numpy as np
from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
import torch

# load all the models and vectorizer (global vocabulary)
Seq_model = load_model("LSTM.h5") # Sequential
SVM_model = joblib.load("SVM_Linear_Kernel.joblib") # SVM
logistic_model = joblib.load("Logistic_Model.joblib") # Logistic
svm_model = joblib.load('svm_model.joblib')

vectorizer = joblib.load("vectorizer.joblib") # global vocabulary (used for Logistic, SVC)
tokenizer = joblib.load("tokenizer.joblib") # used for LSTM

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
tokenizer1 = DistilBertTokenizer.from_pretrained("tokenizer_bert")
model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased', num_labels=5)
model.load_state_dict(torch.load("fine_tuned_bert_model1.pth", map_location=device))

# Decode label function
# {'business': 0, 'entertainment': 1, 'health': 2, 'politics': 3, 'sport': 4}
def decodedLabel(input_number):
    print('receive label encoded', input_number)
    categories = {
      0: 'Business',
      1: 'Entertainment',
      2: 'Health',
      3: 'Politics',
      4: 'Sport'
    }
    result = categories.get(input_number) # Ex: Health
    print('decoded result', result)
    return result

# Web Crawler function
def crawURL(url):
    # Fetch the URL content
    response = requests.get(url)
    # Parse the sitemap HTML
    soup = BeautifulSoup(response.content, 'html.parser')

    # Find all anchor tags that are children of span tags with class 'sitemap-link'
    urls = [span.a['href'] for span in soup.find_all('span', class_='sitemap-link') if span.a]

    # Crawl pages and extract data
    try:
        print(f"Crawling page: {url}")
        # Fetch page content
        page_response = requests.get(url)
        page_content = page_response.content

        # Parse page content with BeautifulSoup
        soup = BeautifulSoup(page_content, 'html.parser')

        # Extract data you need from the page
        author = soup.find("meta", {"name": "author"}).attrs['content'].strip()
        date_published = soup.find("meta", {"property": "article:published_time"}).attrs['content'].strip()
        article_section = soup.find("meta", {"name": "meta-section"}).attrs['content']
        url = soup.find("meta", {"property": "og:url"}).attrs['content']
        headline = soup.find("h1", {"data-editable": "headlineText"}).text.strip()
        description = soup.find("meta", {"name": "description"}).attrs['content'].strip()
        keywords = soup.find("meta", {"name": "keywords"}).attrs['content'].strip()
        text = soup.find(itemprop="articleBody")
        # Find all <p> tags with class "paragraph inline-placeholder"
        paragraphs = text.find_all('p', class_="paragraph inline-placeholder")

        # Initialize an empty list to store the text content of each paragraph
        paragraph_texts = []

        # Iterate over each <p> tag and extract its text content
        for paragraph in paragraphs:
            paragraph_texts.append(paragraph.text.strip())

        # Join the text content of all paragraphs into a single string
        full_text = ''.join(paragraph_texts)
        return full_text
        
    except Exception as e:
        print(f"Failed to crawl page: {url}, Error: {str(e)}")
        return None

# Predict for text category by Models
def process_api(text):
    # Vectorize the text data
    processed_text = vectorizer.transform([text])
    sequence = tokenizer.texts_to_sequences([text])
    padded_sequence = pad_sequences(sequence, maxlen=1000, padding='post')

    new_encoding = tokenizer1([text], truncation=True, padding=True, return_tensors="pt")
    input_ids = new_encoding['input_ids']
    attention_mask = new_encoding['attention_mask']
    with torch.no_grad():
        output = model(input_ids, attention_mask=attention_mask)
        logits = output.logits
    
    # Get the predicted result from models
    Logistic_Predicted = logistic_model.predict(processed_text).tolist() # Logistic Model
    SVM_Predicted = SVM_model.predict(processed_text).tolist() # SVC Model
    Seq_Predicted = Seq_model.predict(padded_sequence)
    predicted_label_index = np.argmax(Seq_Predicted)
    
    # ----------- Proba -----------
    Logistic_Predicted_proba = logistic_model.predict_proba(processed_text)
    svm_new_probs = SVM_model.decision_function(processed_text)
    svm_probs = svm_model.predict_proba(svm_new_probs)
    predicted_label_index = np.argmax(Seq_Predicted)

    bert_probabilities = torch.softmax(logits, dim=1)
    max_probability = torch.max(bert_probabilities).item()
    predicted_label_bert = torch.argmax(logits, dim=1).item()
    # ----------- Debug Logs -----------
    logistic_debug = decodedLabel(int(Logistic_Predicted[0]))
    svc_debug = decodedLabel(int(SVM_Predicted[0]))
    # predicted_label_index = np.argmax(Seq_Predicted)
    #print('Logistic', int(Logistic_Predicted[0]), logistic_debug)
    #print('SVM', int(SVM_Predicted[0]), svc_debug)
    
    return {
            'predicted_label_logistic': decodedLabel(int(Logistic_Predicted[0])),
            'probability_logistic': f"{int(float(np.max(Logistic_Predicted_proba))*10000//100)}%",
        
            'predicted_label_svm': decodedLabel(int(SVM_Predicted[0])),
            'probability_svm': f"{int(float(np.max(svm_probs))*10000//100)}%",
        
            'predicted_label_lstm': decodedLabel(int(predicted_label_index)),
            'probability_lstm': f"{int(float(np.max(Seq_Predicted))*10000//100)}%",

            'predicted_label_bert': decodedLabel(int(predicted_label_bert)),
            'probability_bert': f"{int(float(max_probability)*10000//100)}%",
        
            'Article_Content': text
        }

# Init web crawling, process article content by Model and return result as JSON
def categorize(url):
    try:
        article_content = crawURL(url)
        result = process_api(article_content)
        return result
    except Exception as error:
        if hasattr(error, 'message'):
            return {"error_message": error.message}
        else:
            return {"error_message": error}

        
# Main App
st.title('Instant Category Classification')
st.write("Unsure what category a CNN article belongs to? Our clever tool can help! Paste the URL below and press Enter. We'll sort it into one of our 5 categories in a flash! ⚡️")

# Define category information (modify content and bullet points as needed)
categories = {
    "Business": [
        "Analyze market trends and investment opportunities.",
        "Gain insights into company performance and industry news.",
        "Stay informed about economic developments and regulations."
    ],
    "Health": [
        "Discover healthy recipes and exercise tips.",
        "Learn about the latest medical research and advancements.",
        "Find resources for managing chronic conditions and improving well-being."
    ],
    "Sport": [
        "Follow your favorite sports teams and athletes.",
        "Explore news and analysis from various sports categories.",
        "Stay updated on upcoming games and competitions."
    ],
    "Politics": [
        "Get informed about current political events and policies.",
        "Understand different perspectives on political issues.",
        "Engage in discussions and debates about politics."
    ],
    "Entertainment": [
        "Find recommendations for movies, TV shows, and music.",
        "Explore reviews and insights from entertainment critics.",
        "Stay updated on celebrity news and cultural trends."
    ]
}

# Define model information (modify descriptions as needed)
models = {
  "Logistic Regression": "A widely used statistical method for classification problems. It excels at identifying linear relationships between features and the target variable.",
  "SVC (Support Vector Classifier)": "A powerful machine learning model that seeks to find a hyperplane that best separates data points of different classes. It's effective for high-dimensional data and can handle some non-linear relationships.",
  "LSTM (Long Short-Term Memory)": "A type of recurrent neural network (RNN) particularly well-suited for sequential data like text or time series. LSTMs can effectively capture long-term dependencies within the data.",
  "BERT (Bidirectional Encoder Representations from Transformers)": "A powerful pre-trained model based on the Transformer architecture. It excels at understanding the nuances of language and can be fine-tuned for various NLP tasks like text classification."
}


# CNN URL Example List
URL_Example = [
    'https://edition.cnn.com/2012/01/31/health/frank-njenga-mental-health/index.html',
    'https://edition.cnn.com/2024/04/30/entertainment/barbra-streisand-melissa-mccarthy-ozempic/index.html',
    'https://edition.cnn.com/2024/04/30/sport/lebron-james-lakers-future-nba-spt-intl/index.html',
    'https://edition.cnn.com/2024/04/30/business/us-home-prices-rose-in-february/index.html'
]

# Create expanders containing list of categories can be classified
with st.expander("Category List"):
  # Title for each category
  st.subheader("Available Categories:")
  for category in categories.keys():
    st.write(f"- {category}")
  # Content for each category (separated by a horizontal line)
  st.write("---")
  for category, content in categories.items():
    st.subheader(category)
    for item in content:
      st.write(f"- {item}")


# Create expanders containing list of models used in this project
with st.expander("Available Models"):
  st.subheader("List of Models:")
  for model_name in models.keys():
    st.write(f"- {model_name}")
  st.write("---")
  for model_name, description in models.items():
    st.subheader(model_name)
    st.write(description)

with st.expander("URLs Example"):
  for url in URL_Example:
    st.write(f"- {url}")
      
# Explain to user why this project is only worked for CNN domain
with st.expander("Tips", expanded=True):
    st.write(
        '''
            This project works best with CNN articles right now. 
            Our web crawler is like a special tool for CNN's website. 
            It can't quite understand other websites because they're built differently
        '''
    )

st.divider() # 👈 Draws a horizontal rule

st.title('Dive in! See what category your CNN story belongs to 😉.')
# Paste URL Input
url = st.text_input("Find your favorite CNN story! Paste the URL and press ENTER 🔍.", placeholder='Ex: https://edition.cnn.com/2012/01/31/health/frank-njenga-mental-health/index.html')

if url:
    st.divider() # 👈 Draws a horizontal rule
    result = categorize(url)
    article_content = result.get('Article_Content')
    st.title('Article Content Fetched')
    st.text_area("", value=article_content, height=400) # render the article content as textarea element
    st.divider()  # 👈 Draws a horizontal rule
    st.title('Predicted Results')
    st.json({
        "Logistic": {
            "predicted_label": result.get("predicted_label_logistic"),
            "probability": result.get("probability_logistic")
        },
        "SVC": {
            "predicted_label": result.get("predicted_label_svm"),
            "probability": result.get("probability_svm")
        },
        "LSTM": {
            "predicted_label": result.get("predicted_label_lstm"),
            "probability": result.get("probability_lstm")
        },
        "BERT": {
            "predicted_label": result.get("predicted_label_bert"),
            "probability": result.get("probability_bert")
        }
    })
    
st.divider()  # 👈 Draws a horizontal rule

# Category labels and corresponding counts
categories = ["Sport", "Health", "Entertainment", "Politics", "Business"]
counts = [5638, 4547, 2658, 2461, 1362]

# Optional: Add a chart title
st.title("Training Data Category Distribution")

# Optional: Display additional information
st.write("Here's a breakdown of the number of articles in each category:")
for category, count in zip(categories, counts):
  st.write(f"- {category}: {count}")

# Create the bar chart
st.bar_chart(data=dict(zip(categories, counts)))

st.divider()  # 👈 Draws a horizontal rule

# ------------ Copyright Section ------------
# Get the current year
current_year = date.today().year
# Format the copyright statement with dynamic year
copyright_text = f"Copyright © {current_year}"
st.title(copyright_text)
author_names = ["Trần Thanh Phước (Mentor)", "Lương Ngọc Phương (Member)", "Trịnh Cẩm Minh (Member)"]
st.write("Meet the minds behind the work!")
for author in author_names:
    if (author == "Trịnh Cẩm Minh (Member)"): st.markdown("- [Trịnh Cẩm Minh (Member)](https://minhct.netlify.app/)")
    else: st.markdown(f"- {author}\n")  # Use f-string for bullet and newline