Phase 2 starts

README.md

@@ -1,10 +1,27 @@
 # Toxicity analysis in social media using BERT
 
+## PHASE 1 - SEM 6 
 ## Meet the team 
 - [Chethan](https://github.com/chethanv-20)
 - [Darshan](https://github.com/darshangn310)
 - [Janav](https://github.com/Janav20)
 - [Harshini](https://github.com/harshinimurugan2004)
 
-### This project is still under development process - check repo after some time 
 > Phase 1 of our project ends on 10-05-2024
+
+
+## PHASE 2 - SEM 7 
+
+### Review 1
+- Worked on model
+- curated euphermism dataset
+- Chat room application for the working model
+
+### Streamlit Interface
+<img src="https://github.com/user-attachments/assets/19de66cb-3519-46e7-8213-496ed2e736ad" width="800" height="400">
+
+### warning and blocking 
+<img src="https://github.com/user-attachments/assets/191c39d8-3109-4ece-82cc-b8aa7cf8ed34" width="300" height="200">
+<img src="https://github.com/user-attachments/assets/cb590165-4816-4336-883f-ab1dc73a2e5a" width="300" height="200">
+
+### This project is still under development process - check repo after some time 

app.py

@@ -0,0 +1,146 @@
+import streamlit as st
+from collections import defaultdict
+import pandas as pd
+from model import predict_toxicity_with_context
+
+# Load euphemism dictionary from CSV
+toxic_words_path = 'toxic_words_with_variations.csv'
+toxic_words_df = pd.read_csv(toxic_words_path)
+
+# Flatten the toxic words variations into a dictionary
+euphemism_dict = {}
+for _, row in toxic_words_df.iterrows():
+    base_word = row['Original Word']
+    for col in toxic_words_df.columns[1:]:
+        variation = row[col]
+        if pd.notna(variation):
+            euphemism_dict[variation] = base_word
+
+# Initialize session state variables if they don't exist
+if 'chat_history' not in st.session_state:
+    st.session_state.chat_history = []
+if 'user_toxic_count' not in st.session_state:
+    st.session_state.user_toxic_count = defaultdict(int)
+if 'flagged_users' not in st.session_state:
+    st.session_state.flagged_users = set()
+if 'blocked_users' not in st.session_state:
+    st.session_state.blocked_users = set()
+
+# Function to display chat history with color-coded results
+def display_chat(chat_history):
+    for user, msg, result, is_toxic in chat_history:
+        color = "red" if is_toxic else "green"
+        st.markdown(f"**{user}:** {msg} - <span style='color: {color};'>{result}</span>", unsafe_allow_html=True)
+
+def view_toxicity_report():
+    report = pd.DataFrame(list(st.session_state.user_toxic_count.items()), columns=['User', 'Toxic Message Count'])
+    st.write(report)
+
+# Design and styling
+st.set_page_config(page_title="Toxicity Analysis Chat Room", page_icon=":speech_balloon:", layout="wide")
+st.markdown("<h1 style='text-align: center; color: #00FFFF;'>Toxicity Analysis in Social Media</h1>", unsafe_allow_html=True)
+
+# Define columns for user input to align them horizontally
+col1, col2, col3 = st.columns([1, 1, 1])
+
+def handle_user_input(user, user_input):
+    if user in st.session_state.blocked_users:
+        st.error(f"{user} is blocked and cannot send messages.")
+        return "Blocked", False
+
+    if user_input.strip() == "":  # Check if input is empty
+        st.error("Input cannot be empty!")
+        return "No input provided", False
+
+    result = predict_toxicity_with_context(user_input, euphemism_dict)
+
+    # Debugging: Show the result for inspection
+    st.write(f"Debug: Result from toxicity analysis - '{result}'")
+
+    # Determine if the result indicates toxicity
+    if "classified as non-toxic" in result:
+        is_toxic = False
+    elif "classified as toxic" in result:
+        is_toxic = True
+    else:
+        # Handle unexpected result formats
+        is_toxic = "toxic" in result.lower()
+
+    # Debugging: Show whether the message is classified as toxic
+    st.write(f"Debug: Is toxic - {is_toxic}")
+
+    st.session_state.chat_history.append((user, user_input, result, is_toxic))
+
+    # Update the count only if the message is toxic
+    if is_toxic:
+        st.session_state.user_toxic_count[user] += 1
+        toxic_count = st.session_state.user_toxic_count[user]
+
+        if toxic_count == 3:
+            st.warning(f"{user}, you have sent 3 toxic messages. Please be mindful of your behavior.")
+        elif toxic_count == 5:
+            st.warning(f"{user}, you have sent 5 toxic messages. Further toxic behavior will result in being blocked.")
+        elif toxic_count >= 6:
+            st.session_state.blocked_users.add(user)
+            st.warning(f"{user} has been blocked for repeated toxic behavior. you cannot send messages anymore.")
+
+    return result, is_toxic
+
+with col1:
+    user1_input = st.text_input("User 1:", key="user1_input")
+    if st.button("Send as User 1", key="send_user1"):
+        result, is_toxic = handle_user_input("User 1", user1_input)
+
+with col2:
+    user2_input = st.text_input("User 2:", key="user2_input")
+    if st.button("Send as User 2", key="send_user2"):
+        result, is_toxic = handle_user_input("User 2", user2_input)
+
+with col3:
+    user3_input = st.text_input("User 3:", key="user3_input")
+    if st.button("Send as User 3", key="send_user3"):
+        result, is_toxic = handle_user_input("User 3", user3_input)
+
+# Display chat history
+st.markdown("<hr style='border:1px solid #00FFFF;'>", unsafe_allow_html=True)
+st.markdown("<h3 style='text-align: center; color: #00FFFF;'>Chat History</h3>", unsafe_allow_html=True)
+display_chat(st.session_state.chat_history)
+
+# Button to view toxicity report
+if st.button("View Toxicity Report"):
+    view_toxicity_report()
+
+# Check for flagged users
+if st.session_state.flagged_users:
+    st.markdown("<hr style='border:1px solid red;'>", unsafe_allow_html=True)
+    st.markdown("<h3 style='color: red;'>Flagged Users</h3>", unsafe_allow_html=True)
+    for user in st.session_state.flagged_users:
+        st.warning(f"{user} has been flagged for repeated toxic behavior.")
+
+# Add some background styling
+st.markdown(
+    """
+    <style>
+    body {
+        background-color: #F0F8FF;
+        color: #808080;
+        font-family: Arial, sans-serif;
+    }
+    .stTextInput>div>input {
+        border-radius: 12px;
+    }
+    div.stButton > button {
+        background-color: #000000;
+        color: white;
+        border-radius: 12px;
+    }
+    div.stButton > button:hover {
+        background-color: #333333;
+    }
+    .stMarkdown {
+        font-family: 'Arial', sans-serif;
+    }
+    </style>
+    """,
+    unsafe_allow_html=True
+)

model.py

@@ -0,0 +1,73 @@
+import torch
+from transformers import BertTokenizer, BertForSequenceClassification
+from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
+
+# Load the trained model and tokenizer
+model = BertForSequenceClassification.from_pretrained('./saved_model')
+tokenizer = BertTokenizer.from_pretrained('./saved_model')
+
+# Initialize sentiment analyzer
+sentiment_analyzer = SentimentIntensityAnalyzer()
+
+# Preprocessing function
+def preprocess_input(text, tokenizer, euphemism_dict, max_len):
+    # Replace euphemisms with the base toxic word
+    for euphemism, base_word in euphemism_dict.items():
+        text = text.replace(euphemism, base_word)
+
+    # Tokenize the text
+    encoding = tokenizer.encode_plus(
+        text,
+        add_special_tokens=True,
+        max_length=max_len,
+        return_token_type_ids=False,
+        padding='max_length',
+        truncation=True,
+        return_attention_mask=True,
+        return_tensors='pt',
+    )
+
+    return encoding
+
+# Prediction function with context analysis
+def predict_toxicity_with_context(text, euphemism_dict, max_len=128):
+    model.eval()
+    original_text = text
+    reasons = []
+
+    # Replace euphemisms with the base toxic word
+    for euphemism, base_word in euphemism_dict.items():
+        if euphemism in text:
+            reasons.append(f"Euphemism '{euphemism}' detected, replaced with '{base_word}'")
+            text = text.replace(euphemism, base_word)
+
+    encoding = preprocess_input(text, tokenizer, euphemism_dict, max_len)
+    input_ids = encoding['input_ids']
+    attention_mask = encoding['attention_mask']
+
+    with torch.no_grad():
+        outputs = model(input_ids=input_ids, attention_mask=attention_mask)
+    logits = outputs.logits
+    predictions = torch.sigmoid(logits).cpu().numpy().flatten()
+
+    # Apply sentiment analysis
+    sentiment_score = sentiment_analyzer.polarity_scores(text)['compound']
+
+    # Determine if the text is non-toxic based on positive sentiment
+    if sentiment_score >= 0:
+        return "The input text is classified as non-toxic based on sentiment analysis."
+
+    # Determine toxicity based on threshold
+    threshold = 0.5
+    labels = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']
+    detected_labels = [labels[i] for i in range(len(predictions)) if predictions[i] >= threshold]
+
+    result = ""
+    if detected_labels:
+        result += f"Detected toxicity: {', '.join(detected_labels)}"
+        if reasons:
+            result += f"\nReasons for toxicity: {', '.join(reasons)}"
+    else:
+        result = "The input text is classified as non-toxic."
+
+    return result