File size: 56,606 Bytes
c48986c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 |
import streamlit as st
import pandas as pd
import numpy as np
import re
import matplotlib.pyplot as plt
import seaborn as sns
import nltk
from nltk.corpus import stopwords
from nltk.stem.snowball import SnowballStemmer
import pickle
from transformers import pipeline as hf_pipeline
from sklearn.utils.multiclass import type_of_target
import io
import base64
from sklearn.model_selection import train_test_split
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.pipeline import Pipeline
from sklearn.multiclass import OneVsRestClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.naive_bayes import MultinomialNB
from sklearn.metrics import roc_auc_score, accuracy_score, classification_report
from textblob import TextBlob
import warnings
from sklearn.metrics import precision_score, recall_score, f1_score, confusion_matrix
from sklearn.metrics import roc_curve
warnings.filterwarnings('ignore')
# Download required NLTK resources
try:
nltk.data.find('corpora/stopwords')
except LookupError:
nltk.download('stopwords')
# Initialize the stemmer
stemmer = SnowballStemmer('english')
stop_words_set = set(stopwords.words('english'))
# Text preprocessing functions
def remove_stopwords(text):
return " ".join([word for word in str(text).split() if word.lower() not in stop_words_set])
def train_lightweight_model(data, text_column, label_column):
"""
Train a lightweight model for toxicity detection
Args:
data: DataFrame containing the data
text_column: Column name for text data
label_column: Column name for label data
Returns:
Trained model and vectorizer
"""
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import Pipeline
# Preprocess text
data['processed_text'] = data[text_column].apply(preprocess_text)
# Create a pipeline with TF-IDF and Logistic Regression
model = Pipeline([
('tfidf', TfidfVectorizer(max_features=5000, ngram_range=(1, 2))),
('clf', LogisticRegression(random_state=42, max_iter=1000))
])
# Train the model
model.fit(data['processed_text'], data[label_column])
return model
def load_bert_model():
"""
Load a pre-trained BERT model for sentiment analysis
Returns:
Loaded model
"""
try:
# Load sentiment analysis model from HuggingFace
sentiment_analyzer = hf_pipeline("sentiment-analysis")
st.success("BERT model loaded successfully!")
return sentiment_analyzer
except Exception as e:
st.error(f"Error loading BERT model: {e}")
return None
def clean_text(text):
text = str(text).lower()
text = re.sub(r"what's", "what is ", text)
text = re.sub(r"\'s", " ", text)
text = re.sub(r"\'ve", " have ", text)
text = re.sub(r"can't", "can not ", text)
text = re.sub(r"n't", " not ", text)
text = re.sub(r"i'm", "i am ", text)
text = re.sub(r"\'re", " are ", text)
text = re.sub(r"\'d", " would ", text)
text = re.sub(r"\'ll", " will ", text)
text = re.sub(r"\'scuse", " excuse ", text)
text = re.sub(r'\W', ' ', text) # Remove non-word characters
text = re.sub(r'\s+', ' ', text).strip() # Remove extra spaces
return text
def stemming(sentence):
return " ".join([stemmer.stem(word) for word in str(sentence).split()])
def preprocess_text(text):
text = remove_stopwords(text)
text = clean_text(text)
text = stemming(text)
return text
# Function to get sentiment
def get_sentiment(text):
score = TextBlob(text).sentiment.polarity
if score > 0:
return "Positive", score
elif score < 0:
return "Negative", score
else:
return "Neutral", score
# Function to moderate text based on toxicity
def moderate_text(text, predictions, threshold_moderate=0.5, threshold_delete=0.8):
# If binary classification, only check the 'toxic' probability (index 1)
if len(predictions) == 2:
toxic_prob = predictions[1]
if toxic_prob >= threshold_delete:
return "*** COMMENT DELETED DUE TO HIGH TOXICITY ***", "delete"
elif toxic_prob >= threshold_moderate:
# List of potentially toxic words to censor
toxic_words = ["stupid", "idiot", "dumb", "hate", "sucks", "terrible",
"awful", "garbage", "trash", "pathetic", "ridiculous"]
words = text.split()
moderated_words = []
for word in words:
# Clean word for comparison
clean_word = re.sub(r'[^\w\s]', '', word.lower())
# Check if the word is in the toxic words list
if clean_word in toxic_words:
# Replace with a more neutral placeholder
moderated_words.append("[inappropriate]")
else:
moderated_words.append(word)
return " ".join(moderated_words), "moderate"
else:
return text, "keep"
else:
# Multi-label: check all classes
if any(pred >= threshold_delete for pred in predictions):
return "*** COMMENT DELETED DUE TO HIGH TOXICITY ***", "delete"
elif any(pred >= threshold_moderate for pred in predictions):
# List of potentially toxic words to censor
toxic_words = ["stupid", "idiot", "dumb", "hate", "sucks", "terrible",
"awful", "garbage", "trash", "pathetic", "ridiculous"]
words = text.split()
moderated_words = []
for word in words:
# Clean word for comparison
clean_word = re.sub(r'[^\w\s]', '', word.lower())
# Check if the word is in the toxic words list
if clean_word in toxic_words:
# Replace with a more neutral placeholder
moderated_words.append("[inappropriate]")
else:
moderated_words.append(word)
return " ".join(moderated_words), "moderate"
else:
return text, "keep"
# Function to train and save the model
def train_model(X_train, y_train, model_type='logistic_regression'):
st.write("Training model...")
# Ensure `y_train` has 6 columns
label_columns = ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']
# Create missing columns if they don't exist
for col in label_columns:
if col not in y_train.columns:
y_train[col] = 0
# Ensure columns are in the right order
y_train = y_train[label_columns]
if model_type == 'logistic_regression':
pipeline = Pipeline([
('tfidf', TfidfVectorizer(stop_words='english', max_features=50000)),
('clf', OneVsRestClassifier(LogisticRegression(max_iter=1000), n_jobs=-1))
])
else: # Naive Bayes
pipeline = Pipeline([
('tfidf', TfidfVectorizer(stop_words='english', max_features=50000)),
('clf', OneVsRestClassifier(MultinomialNB(), n_jobs=-1))
])
pipeline.fit(X_train, y_train)
return pipeline
def evaluate_model(pipeline, X_test, y_test):
"""
Evaluates the given trained pipeline on test data.
Returns:
accuracy: Accuracy score
roc_auc: ROC AUC score
predictions: Predicted labels
pred_probs: Predicted probabilities
fpr: False Positive Rate array (for ROC curve)
tpr: True Positive Rate array (for ROC curve)
"""
# Get predictions and prediction probabilities
predictions = pipeline.predict(X_test)
pred_probs = pipeline.predict_proba(X_test)
if isinstance(pred_probs, list) and len(pred_probs) == 1:
pred_probs = pred_probs[0] # Handle list with 1 array
# Ensure predictions format matches y_test
y_type = type_of_target(y_test)
pred_type = type_of_target(predictions)
if y_type != pred_type:
if y_type == "multilabel-indicator" and pred_type == "binary":
# Expand binary predictions to multilabel shape
predictions = np.array([[pred] * y_test.shape[1] for pred in predictions])
elif y_type == "binary" and pred_type == "multilabel-indicator":
# Collapse multilabel predictions to binary
predictions = predictions[:, 0]
# Calculate accuracy
accuracy = accuracy_score(y_test, predictions)
# Calculate ROC AUC
try:
if len(y_test.shape) > 1 and y_test.shape[1] > 1:
# Multi-label case
roc_auc_sum = 0
valid_labels = 0
for i in range(y_test.shape[1]):
try:
roc_auc_sum += roc_auc_score(y_test.iloc[:, i], pred_probs[:, i])
valid_labels += 1
except Exception:
continue
roc_auc = roc_auc_sum / valid_labels if valid_labels > 0 else 0.0
else:
# Binary case
roc_auc = roc_auc_score(y_test, pred_probs[:, 1] if pred_probs.ndim > 1 and pred_probs.shape[1] > 1 else pred_probs)
except Exception as e:
print(f"Warning: Could not compute ROC AUC - {e}")
roc_auc = 0.0
# Calculate FPR, TPR for ROC curve (only for binary classification)
try:
if len(y_test.shape) == 1 or (len(y_test.shape) > 1 and y_test.shape[1] == 1):
fpr, tpr, _ = roc_curve(
y_test,
pred_probs[:, 1] if pred_probs.ndim > 1 and pred_probs.shape[1] > 1 else pred_probs
)
else:
fpr, tpr = None, None # ROC Curve not available for multilabel
except Exception as e:
print(f"Warning: Could not compute ROC Curve - {e}")
fpr, tpr = None, None
return accuracy, roc_auc, predictions, pred_probs, fpr, tpr
# Function to create a download link for the trained model
def get_model_download_link(model, filename):
model_bytes = pickle.dumps(model)
b64 = base64.b64encode(model_bytes).decode()
href = f'<a href="data:file/pickle;base64,{b64}" download="{filename}">Download Trained Model</a>'
return href
# Function to plot toxicity distribution
def plot_toxicity_distribution(df, toxicity_columns):
fig, ax = plt.subplots(figsize=(10, 6))
x = df[toxicity_columns].sum()
sns.barplot(x=x.index, y=x.values, alpha=0.8, palette='viridis', ax=ax)
plt.title('Toxicity Distribution')
plt.ylabel('Count')
plt.xlabel('Toxicity Category')
plt.xticks(rotation=45)
return fig
# Function to provide sample data format
def show_sample_data_format():
st.subheader("Sample Data Format")
# Create sample dataframe
sample_data = {
'comment_text': [
"This is a normal comment.",
"This is a toxic comment you idiot!",
"You're all worthless and should die.",
"I respectfully disagree with your point."
],
'toxic': [0, 1, 1, 0],
'severe_toxic': [0, 0, 1, 0],
'obscene': [0, 1, 0, 0],
'threat': [0, 0, 1, 0],
'insult': [0, 1, 1, 0],
'identity_hate': [0, 0, 0, 0]
}
sample_df = pd.DataFrame(sample_data)
st.dataframe(sample_df)
# Create download link for sample data
csv = sample_df.to_csv(index=False)
b64 = base64.b64encode(csv.encode()).decode()
href = f'<a href="data:file/csv;base64,{b64}" download="sample_toxic_data.csv">Download Sample CSV</a>'
st.markdown(href, unsafe_allow_html=True)
st.info("""
Your CSV file should contain:
1. A column with comment text
2. One or more columns with binary values (0 or 1) for each toxicity category
""")
# Function to validate dataset
def validate_dataset(df, comment_column, toxicity_columns):
issues = []
# Check if comment column exists
if comment_column not in df.columns:
issues.append(f"Comment column '{comment_column}' not found in the dataset")
# Check if toxicity columns exist
missing_columns = [col for col in toxicity_columns if col not in df.columns]
if missing_columns:
issues.append(f"Missing toxicity columns: {', '.join(missing_columns)}")
# Check if values in toxicity columns are valid (0 or 1)
for col in toxicity_columns:
if col in df.columns:
# Check for non-numeric values
if not pd.api.types.is_numeric_dtype(df[col]):
issues.append(f"Column '{col}' contains non-numeric values")
else:
# Check for values other than 0 and 1
invalid_values = df[col].dropna().apply(lambda x: x not in [0, 1, 0.0, 1.0])
if invalid_values.any():
issues.append(f"Column '{col}' contains values other than 0 and 1")
# Check for empty data
if df.empty:
issues.append("Dataset is empty")
elif df[comment_column].isna().all():
issues.append("Comment column contains no data")
return issues
# Function to extract predictions from model output
def extract_predictions(predictions_proba, toxicity_categories):
"""
Helper function to extract probabilities from model output,
handling different output formats.
"""
# Debug information
if st.session_state.debug_mode:
st.write(f"Predictions type: {type(predictions_proba)}")
st.write(
f"Predictions shape/length: {np.shape(predictions_proba) if hasattr(predictions_proba, 'shape') else len(predictions_proba)}")
# Case 1: List of arrays with one element per toxicity category
if isinstance(predictions_proba, list) and len(predictions_proba) == len(toxicity_categories):
return [pred_array[:, 1][0] if pred_array.shape[1] > 1 else pred_array[0] for pred_array in predictions_proba]
# Case 2: List with a single array (common for OneVsRestClassifier)
elif isinstance(predictions_proba, list) and len(predictions_proba) == 1:
pred_array = predictions_proba[0]
# If it's a 2D array with number of columns equal to number of categories
if len(pred_array.shape) == 2 and pred_array.shape[1] == len(toxicity_categories):
return pred_array[0] # Return first row, which contains all probabilities
# If it's a 2D array with 2 columns per category (common binary classifier output)
elif len(pred_array.shape) == 2 and pred_array.shape[1] == 2:
return np.array([pred_array[0, 1]])
# Case 3: Direct numpy array
elif isinstance(predictions_proba, np.ndarray):
# If it's already the right shape
if len(predictions_proba.shape) == 2 and predictions_proba.shape[1] == len(toxicity_categories):
return predictions_proba[0]
# If it's a 2D array with two columns (binary classification)
elif len(predictions_proba.shape) == 2 and predictions_proba.shape[1] == 2:
# For binary classification, return the probability of positive class
return np.array([predictions_proba[0, 1]])
# If prediction format isn't recognized, return a repeated array of single probability
# This handles the case where we only have one prediction but need to repeat it
if isinstance(predictions_proba, list) and len(predictions_proba) == 1:
single_prob = predictions_proba[0]
if hasattr(single_prob, 'shape') and len(single_prob.shape) == 2 and single_prob.shape[1] == 2:
# Take positive class probability and repeat for all categories
return np.full(len(toxicity_categories), single_prob[0, 1])
# Last resort fallback
st.warning(f"Unexpected prediction format. Creating default predictions.")
return np.zeros(len(toxicity_categories))
def display_classification_result(result):
st.subheader("Classification Result")
# Show original and moderated text side by side
col1, col2 = st.columns(2)
with col1:
st.markdown("**Original Text**")
st.code(result["original_text"], language="text")
with col2:
st.markdown("**Moderated Text**")
st.code(result["moderated_text"], language="text")
# Show action with color
action = result["action"]
if action == "keep":
st.success("✅ This comment is allowed (Non-toxic).")
elif action == "moderate":
st.warning("⚠️ This comment is moderated (Potentially toxic).")
elif action == "delete":
st.error("🚫 This comment is deleted (Highly toxic).")
# Show toxicity scores
st.markdown("**Toxicity Scores:**")
score_cols = st.columns(len(result["toxicity_scores"]))
for i, (label, score) in enumerate(result["toxicity_scores"].items()):
score_cols[i].metric(label.capitalize(), f"{score:.2%}")
# Show sentiment if available
if "sentiment" in result:
st.markdown("**Sentiment Analysis:**")
st.info(f"{result['sentiment']['label']} (score: {result['sentiment']['score']:.2%})")
def moderate_comment(comment, model, sentiment_model=None):
"""
Moderate a single comment using the trained model and optionally BERT sentiment analysis.
Args:
comment: The comment text to moderate
model: The trained model to use for toxicity detection
sentiment_model: Optional BERT model for sentiment analysis
Returns:
Dictionary containing moderation results
"""
# Preprocess the comment
processed_text = preprocess_text(comment)
# Get model predictions
predictions = model.predict_proba([processed_text])[0]
# Get sentiment if BERT model is available
sentiment = None
if sentiment_model:
sentiment = sentiment_model(comment)[0]
# Moderate the text
moderated_text, action = moderate_text(comment, predictions)
# Prepare result
result = {
"original_text": comment,
"moderated_text": moderated_text,
"action": action,
"toxicity_scores": {}
}
# Handle both binary and multi-class predictions
if len(predictions) == 2: # Binary classification
result["toxicity_scores"] = {
"toxic": float(predictions[1]), # Probability of positive class
"non_toxic": float(predictions[0]) # Probability of negative class
}
else: # Multi-class classification
# Define the toxicity categories
categories = ["toxic", "severe_toxic", "obscene", "threat", "insult", "identity_hate"]
# Add scores for each category that exists in the predictions
for i, category in enumerate(categories):
if i < len(predictions):
result["toxicity_scores"][category] = float(predictions[i])
if sentiment:
result["sentiment"] = {
"label": sentiment["label"],
"score": float(sentiment["score"])
}
return result
# --- Bias Detection Module ---
def detect_subgroup(text):
gender_keywords = ["he", "she", "him", "her", "man", "woman", "boy", "girl", "male", "female"]
ethnicity_keywords = [
"asian", "black", "white", "hispanic", "latino", "indian", "african", "european", "arab", "jewish", "muslim", "christian"
]
text_lower = text.lower()
subgroups = set()
if any(word in text_lower for word in gender_keywords):
subgroups.add("gender")
if any(word in text_lower for word in ethnicity_keywords):
subgroups.add("ethnicity")
return list(subgroups)
def bias_report(X, y_true, y_pred, text_column_name):
# X: DataFrame with text column
# y_pred: predicted labels (same shape as y_true)
# y_true: true labels (same shape as y_pred)
# text_column_name: name of the text column
results = []
for idx, row in X.iterrows():
subgroups = detect_subgroup(row[text_column_name])
if subgroups:
for subgroup in subgroups:
results.append({
"subgroup": subgroup,
"is_toxic": int(y_pred[idx].sum() > 0) if len(y_pred.shape) > 1 else int(y_pred[idx] > 0)
})
if not results:
return "No sensitive subgroups detected in the evaluation set."
df = pd.DataFrame(results)
report = ""
for subgroup in df["subgroup"].unique():
total = (df["subgroup"] == subgroup).sum()
toxic = df[(df["subgroup"] == subgroup) & (df["is_toxic"] == 1)].shape[0]
rate = toxic / total if total > 0 else 0
report += f"- **{subgroup.capitalize()}**: {toxic}/{total} ({rate:.1%}) flagged as toxic\n"
return report
# Streamlit app
def main():
st.set_page_config(
page_title="Toxic Comment Classifier",
page_icon="🧊",
layout="wide",
initial_sidebar_state="expanded",
)
col1, col2 = st.columns([1, 4]) # Adjust the ratio as needed
with col1:
st.image("logo.jpeg", width=100) # Smaller width fits better
with col2:
st.title("Toxic Comment Classifier")
# Initialize session state variables if they don't exist
if 'data' not in st.session_state:
st.session_state.data = None
if 'model' not in st.session_state:
st.session_state.model = None
if 'vectorizer' not in st.session_state:
st.session_state.vectorizer = None
if 'predictions' not in st.session_state:
st.session_state.predictions = None
if 'lightweight_model' not in st.session_state:
st.session_state.lightweight_model = None
if 'bert_model' not in st.session_state:
st.session_state.bert_model = None
# Create sidebar navigation
st.sidebar.title("Navigation")
page = st.sidebar.radio(
"Select a page",
["Home", "Data Preprocessing", "Model Training", "Model Evaluation", "Prediction", "Visualization"]
)
# Home page
if page == "Home":
st.header("Home")
st.write("""
Welcome to the Toxic Comment Classifier application. This tool helps you to:
1. Upload and preprocess data
2. Train a machine learning model to detect toxic comments
3. Evaluate model performance
4. Make predictions on new data
5. Visualize results
Please use the sidebar navigation to get started.
""")
# Add option to load BERT sentiment model
if st.sidebar.checkbox("Use BERT for Sentiment Analysis"):
st.subheader("BERT-Based Sentiment Analysis")
st.write("This option uses a pre-trained BERT model for advanced sentiment analysis.")
if st.button("Load BERT Model"):
with st.spinner("Loading BERT model..."):
st.session_state.bert_model = load_bert_model()
st.write("DEBUG: bert_model in session_state after loading:", st.session_state.bert_model)
# Sample data section
st.subheader("Sample Data Format")
show_sample_data_format()
# Single comment moderation
st.subheader("Try Comment Moderation")
comment = st.text_area("Enter a comment to moderate:")
col1, col2 = st.columns(2)
with col1:
use_default_model = st.checkbox("Use built-in model for demo", value=True)
with col2:
use_bert = st.checkbox("Use BERT model for sentiment (if loaded)", value=False)
if st.button("Moderate Comment"):
if comment:
with st.spinner("Analyzing comment..."):
st.write("DEBUG: bert_model in session_state before use:", st.session_state.bert_model)
sentiment_model = st.session_state.bert_model if use_bert and st.session_state.bert_model is not None else None
if use_default_model or st.session_state.model or st.session_state.lightweight_model:
model_to_use = None
if st.session_state.model:
model_to_use = st.session_state.model
elif st.session_state.lightweight_model:
model_to_use = st.session_state.lightweight_model
result = moderate_comment(comment, model_to_use, sentiment_model)
display_classification_result(result)
else:
st.error("No model available. Please train a model first or enable the demo model.")
else:
st.warning("Please enter a comment to moderate.")
# Data Preprocessing page
elif page == "Data Preprocessing":
st.header("Data Preprocessing")
# File upload
st.subheader("Upload Dataset")
uploaded_file = st.file_uploader("Choose a CSV file", type="csv")
if uploaded_file is not None:
try:
# Load data
data = pd.read_csv(uploaded_file)
# Display raw data
st.subheader("Raw Data")
st.write(data.head())
# Validate the dataset
validation_result = validate_dataset(data, 'comment_text', ['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate'])
if not validation_result: # Empty list means no issues
st.success("Dataset is valid!")
# Store data in session state
st.session_state.data = data
# Data cleaning
st.subheader("Data Cleaning")
st.write("Select columns to include in the analysis:")
# Get all columns
all_columns = data.columns.tolist()
default_selected = ["comment_text", "toxic", "severe_toxic", "obscene", "threat", "insult",
"identity_hate"]
default_selected = [col for col in default_selected if col in all_columns]
selected_columns = st.multiselect(
"Select columns",
options=all_columns,
default=default_selected
)
if selected_columns:
# Filter data by selected columns
filtered_data = data[selected_columns]
# Display cleaned data
st.subheader("Filtered Data")
st.write(filtered_data.head())
# Display data statistics
st.subheader("Data Statistics")
st.write(filtered_data.describe())
# Check for missing values
st.subheader("Missing Values")
missing_values = filtered_data.isnull().sum()
st.write(missing_values)
# Handle missing values if any
if missing_values.sum() > 0:
st.warning("There are missing values in the dataset.")
if st.button("Handle Missing Values"):
# Fill missing text with empty string
text_columns = [col for col in selected_columns if filtered_data[col].dtype == 'object']
for col in text_columns:
filtered_data[col] = filtered_data[col].fillna("")
# Fill missing numerical values with 0
numerical_columns = [col for col in selected_columns if
filtered_data[col].dtype != 'object']
for col in numerical_columns:
filtered_data[col] = filtered_data[col].fillna(0)
st.success("Missing values handled!")
st.write(filtered_data.isnull().sum())
# Text preprocessing
st.subheader("Text Preprocessing")
# Select text column
text_columns = [col for col in selected_columns if filtered_data[col].dtype == 'object']
if text_columns:
text_column = st.selectbox("Select text column for preprocessing", text_columns)
# Show sample of original text
st.write("Sample original text:")
sample_texts = filtered_data[text_column].head(3).tolist()
for i, text in enumerate(sample_texts):
st.text(f"Text {i + 1}: {text[:200]}...")
# Preprocess text
if st.button("Preprocess Text"):
with st.spinner("Preprocessing text..."):
filtered_data['processed_text'] = filtered_data[text_column].apply(preprocess_text)
# Show sample of preprocessed text
st.write("Sample preprocessed text:")
sample_preprocessed = filtered_data['processed_text'].head(3).tolist()
for i, text in enumerate(sample_preprocessed):
st.text(f"Processed Text {i + 1}: {text[:200]}...")
# Store preprocessed data
st.session_state.data = filtered_data
st.success("Text preprocessing completed!")
else:
st.warning("No text columns found in the selected columns.")
else:
st.warning("Please select at least one column.")
else:
st.error(f"Dataset validation failed: {validation_result['reason']}")
st.warning("Please upload a valid dataset.")
except Exception as e:
st.error(f"Error loading data: {e}")
st.warning("Please upload a valid CSV file.")
else:
st.info("Please upload a CSV file to begin preprocessing.")
# Model Training page
elif page == "Model Training":
st.header("Model Training")
# Check if data is available
if st.session_state.data is not None:
# Display data info
st.subheader("Dataset Information")
st.write(f"Number of samples: {len(st.session_state.data)}")
if 'processed_text' in st.session_state.data.columns:
st.write("Text preprocessing: Done")
else:
st.warning("Text preprocessing is not done. Please preprocess the data first.")
# Model training options
st.subheader("Training Options")
# Select target column
numerical_columns = [col for col in st.session_state.data.columns if
st.session_state.data[col].dtype != 'object']
if numerical_columns:
target_column = st.selectbox("Select target column", numerical_columns)
# Set training parameters
st.write("Training Parameters:")
col1, col2 = st.columns(2)
with col1:
test_size = st.slider("Test Size", 0.1, 0.5, 0.2, 0.05)
with col2:
random_state = st.number_input("Random State", 0, 100, 42, 1)
# Model selection
model_type = st.radio(
"Select model type",
["Standard Model", "Lightweight Model"]
)
# Train model button
if st.button("Train Model"):
with st.spinner("Training model..."):
# Check if processed text is available
if 'processed_text' in st.session_state.data.columns:
try:
if model_type == "Standard Model":
# Train standard model
X_train = st.session_state.data['processed_text']
y_train = st.session_state.data[[target_column]]
model = train_model(X_train, y_train, 'logistic_regression')
# Store model in session state
st.session_state.model = model
st.session_state.vectorizer = None # Vectorizer is part of the pipeline
st.success("Model training completed!")
else:
# Train lightweight model
lightweight_model = train_lightweight_model(
st.session_state.data,
'processed_text',
target_column
)
# Store lightweight model in session state
st.session_state.lightweight_model = lightweight_model
st.success("Lightweight model training completed!")
except Exception as e:
st.error(f"Error training model: {e}")
else:
st.error("Processed text not found. Please preprocess the data first.")
else:
st.warning("No numerical columns found in the dataset. Please ensure you have target columns.")
else:
st.info("Please upload and preprocess data before training a model.")
# Model Evaluation page
elif page == "Model Evaluation":
st.header("Model Evaluation")
# Check if model is available
model_available = st.session_state.model is not None or st.session_state.lightweight_model is not None
if model_available:
# Display model info
st.subheader("Model Information")
if st.session_state.model is not None:
st.write("Standard model is trained and ready.")
if st.session_state.lightweight_model is not None:
st.write("Lightweight model is trained and ready.")
# Select model to evaluate
model_choice = None
if st.session_state.model is not None and st.session_state.lightweight_model is not None:
model_choice = st.radio(
"Select model to evaluate",
["Standard Model", "Lightweight Model"]
)
# Model evaluation
st.subheader("Evaluation Options")
# Set evaluation parameters
st.write("Evaluation Parameters:")
col1, col2 = st.columns(2)
with col1:
test_size = st.slider("Test Size (Evaluation)", 0.1, 0.5, 0.2, 0.05)
with col2:
random_state = st.number_input("Random State (Evaluation)", 0, 100, 42, 1)
# Target column selection
if st.session_state.data is not None:
numerical_columns = [col for col in st.session_state.data.columns if
st.session_state.data[col].dtype != 'object']
if numerical_columns:
target_column = st.selectbox("Select target column for evaluation", numerical_columns)
# Evaluate model button
if st.button("Evaluate Model"):
with st.spinner("Evaluating model..."):
try:
# Determine which model to evaluate
model_to_evaluate = None
if model_choice == "Lightweight Model" or (
model_choice is None and st.session_state.model is None):
model_to_evaluate = st.session_state.lightweight_model
else:
model_to_evaluate = st.session_state.model
# 1️⃣ Evaluate model
X_test = st.session_state.data['processed_text']
y_test = st.session_state.data[[target_column]]
accuracy, roc_auc, predictions, pred_probs, fpr, tpr = evaluate_model(model_to_evaluate,
X_test, y_test)
# 2️⃣ Calculate additional metrics
precision = precision_score(y_test, predictions, average='weighted', zero_division=0)
recall = recall_score(y_test, predictions, average='weighted', zero_division=0)
f1 = f1_score(y_test, predictions, average='weighted', zero_division=0)
conf_matrix = confusion_matrix(y_test, predictions)
classification_rep = classification_report(y_test, predictions, zero_division=0)
# 3️⃣ Display evaluation metrics
st.subheader("Evaluation Results")
metrics_df = pd.DataFrame({
'Metric': ['Accuracy', 'Precision', 'Recall', 'F1 Score', 'ROC AUC'],
'Value': [accuracy, precision, recall, f1, roc_auc]
})
st.table(metrics_df)
# 4️⃣ Confusion Matrix
st.subheader("Confusion Matrix")
fig, ax = plt.subplots(figsize=(8, 6))
sns.heatmap(conf_matrix, annot=True, fmt='d', cmap='Blues', ax=ax, cbar=False,
annot_kws={"size": 16})
plt.xlabel('Predicted')
plt.ylabel('Actual')
plt.title('Confusion Matrix')
st.pyplot(fig)
# 5️⃣ ROC Curve
st.subheader("ROC Curve")
if fpr is not None and tpr is not None:
fig, ax = plt.subplots(figsize=(8, 6))
ax.plot(fpr, tpr, label=f'ROC Curve (AUC = {roc_auc:.2f})')
ax.plot([0, 1], [0, 1], 'k--')
ax.set_xlabel('False Positive Rate')
ax.set_ylabel('True Positive Rate')
ax.set_title('ROC Curve')
ax.legend(loc="lower right")
st.pyplot(fig)
else:
st.info("ROC curve is not available for multi-label classification.")
# 6️⃣ Classification Report
st.subheader("Classification Report")
st.text(classification_rep)
except Exception as e:
st.error(f"Error evaluating model: {e}")
# 6️⃣ Classification Report
st.subheader("Classification Report")
st.text(classification_rep)
# ✅ Always show bias detection report if possible
# Bias Detection Report
st.subheader("Bias Detection Report")
if 'comment_text' in st.session_state.data.columns:
bias_summary = bias_report(
st.session_state.data[["comment_text"]].reset_index(drop=True),
y_test.reset_index(drop=True),
predictions,
"comment_text" # Add the text_column_name parameter
)
st.markdown(bias_summary)
else:
st.info("No comment_text column found for bias analysis.")
except Exception as e:
st.error(f"Error evaluating model: {e}")
else:
st.warning("No numerical columns found in the dataset. Please ensure you have target columns.")
else:
st.warning("Dataset not available. Please upload and preprocess data first.")
# Model download
st.subheader("Model Download")
# Create download button
model_to_download = None
if model_choice == "Lightweight Model" or (model_choice is None and st.session_state.model is None):
model_to_download = st.session_state.lightweight_model
else:
model_to_download = st.session_state.model
if model_to_download is not None:
# Determine the appropriate filename based on model type
filename = "lightweight_model.pkl" if model_choice == "Lightweight Model" or (model_choice is None and st.session_state.model is None) else "standard_model.pkl"
download_link = get_model_download_link(model_to_download, filename)
st.markdown(download_link, unsafe_allow_html=True)
else:
st.info("Please train a model before evaluation.")
# Prediction page
elif page == "Prediction":
st.header("Prediction")
# Check if model is available
model_available = st.session_state.model is not None or st.session_state.lightweight_model is not None
if model_available:
# Display model info
st.subheader("Model Information")
if st.session_state.model is not None:
st.write("Standard model is trained and ready.")
if st.session_state.lightweight_model is not None:
st.write("Lightweight model is trained and ready.")
# Select model to use
model_choice = None
if st.session_state.model is not None and st.session_state.lightweight_model is not None:
model_choice = st.radio(
"Select model for prediction",
["Standard Model", "Lightweight Model"]
)
# Determine which model to use
model_to_use = None
if model_choice == "Lightweight Model" or (model_choice is None and st.session_state.model is None):
model_to_use = st.session_state.lightweight_model
else:
model_to_use = st.session_state.model
# Prediction options
st.subheader("Make Predictions")
prediction_type = st.radio(
"Select prediction type",
["Single Comment", "Multiple Comments"]
)
# Option to use BERT model
use_bert = False
if st.session_state.bert_model is not None:
use_bert = st.checkbox("Include sentiment analysis with BERT")
# Single comment prediction
if prediction_type == "Single Comment":
comment = st.text_area("Enter a comment to classify:")
if st.button("Classify Comment"):
if comment:
with st.spinner("Classifying comment..."):
st.write("DEBUG: bert_model in session_state before use:", st.session_state.bert_model)
sentiment_model = st.session_state.bert_model if use_bert and st.session_state.bert_model is not None else None
result = moderate_comment(comment, model_to_use, sentiment_model)
display_classification_result(result)
else:
st.warning("Please enter a comment to classify.")
# Multiple comments prediction
else:
# File upload for prediction
uploaded_file = st.file_uploader("Upload a CSV file with comments", type="csv")
if uploaded_file is not None:
try:
# Load data
pred_data = pd.read_csv(uploaded_file)
# Display data
st.subheader("Uploaded Data")
st.write(pred_data.head())
# Select text column
text_columns = [col for col in pred_data.columns if pred_data[col].dtype == 'object']
if text_columns:
text_column = st.selectbox("Select text column for prediction", text_columns)
# Batch prediction button
if st.button("Run Batch Prediction"):
with st.spinner("Classifying comments..."):
try:
# Preprocess text
pred_data['processed_text'] = pred_data[text_column].apply(preprocess_text)
# Run prediction
sentiment_model = st.session_state.bert_model if use_bert else None
predictions = extract_predictions(pred_data, text_column, model_to_use,
sentiment_model)
# Store predictions
st.session_state.predictions = predictions
# Display results
st.subheader("Prediction Results")
st.write(predictions.head())
# Summary
st.subheader("Summary")
toxic_count = predictions['is_toxic'].sum()
total_count = len(predictions)
toxic_percentage = (toxic_count / total_count) * 100
st.write(f"Total comments: {total_count}")
st.write(f"Toxic comments: {toxic_count} ({toxic_percentage:.2f}%)")
st.write(
f"Non-toxic comments: {total_count - toxic_count} ({100 - toxic_percentage:.2f}%)")
# Download predictions
if not predictions.empty:
csv = predictions.to_csv(index=False)
b64 = base64.b64encode(csv.encode()).decode()
href = f'<a href="data:file/csv;base64,{b64}" download="predictions.csv">Download Predictions CSV</a>'
st.markdown(href, unsafe_allow_html=True)
except Exception as e:
st.error(f"Error during prediction: {e}")
else:
st.warning("No text columns found in the uploaded file.")
except Exception as e:
st.error(f"Error loading data: {e}")
st.warning("Please upload a valid CSV file.")
else:
st.info("Please upload a CSV file with comments for batch prediction.")
else:
st.info("Please train a model before making predictions.")
# Visualization page
elif page == "Visualization":
st.header("Visualization")
# Check if data is available
if st.session_state.data is not None:
# Data visualization
st.subheader("Data Visualization")
# Select visualization type
viz_type = st.selectbox(
"Select visualization type",
["Toxicity Distribution", "Comment Length Distribution", "Word Cloud", "Correlation Matrix"]
)
# Toxicity Distribution
if viz_type == "Toxicity Distribution":
# Check if there are label columns
label_columns = [col for col in st.session_state.data.columns if col in [
"toxic", "severe_toxic", "obscene", "threat", "insult", "identity_hate"
]]
if label_columns:
st.write("Toxicity Distribution:")
# Plot toxicity distribution
fig = plot_toxicity_distribution(st.session_state.data, label_columns)
st.pyplot(fig)
else:
st.warning("No toxicity label columns found in the dataset.")
# Comment Length Distribution
elif viz_type == "Comment Length Distribution":
# Check if there is text column
text_columns = [col for col in st.session_state.data.columns if
st.session_state.data[col].dtype == 'object']
if text_columns:
text_column = st.selectbox("Select text column", text_columns)
# Calculate comment lengths
st.session_state.data['comment_length'] = st.session_state.data[text_column].apply(
lambda x: len(str(x)))
# Plot comment length distribution
fig, ax = plt.subplots(figsize=(10, 6))
sns.histplot(st.session_state.data['comment_length'], bins=50, kde=True, ax=ax)
plt.xlabel('Comment Length')
plt.ylabel('Frequency')
plt.title('Comment Length Distribution')
st.pyplot(fig)
# Statistics
st.write("Comment Length Statistics:")
st.write(st.session_state.data['comment_length'].describe())
else:
st.warning("No text columns found in the dataset.")
# Word Cloud
elif viz_type == "Word Cloud":
# Check if there is processed text
if 'processed_text' in st.session_state.data.columns:
try:
from wordcloud import WordCloud
# Create word cloud
st.write("Word Cloud Visualization:")
# Combine all processed text
all_text = ' '.join(st.session_state.data['processed_text'].tolist())
# Generate word cloud
wordcloud = WordCloud(
width=800,
height=400,
background_color='white',
max_words=200,
contour_width=3,
contour_color='steelblue'
).generate(all_text)
# Display word cloud
fig, ax = plt.subplots(figsize=(10, 6))
ax.imshow(wordcloud, interpolation='bilinear')
ax.axis('off')
plt.tight_layout()
st.pyplot(fig)
except ImportError:
st.error("WordCloud package is not installed. Please install it to use this feature.")
else:
st.warning("Processed text not found. Please preprocess the data first.")
# Correlation Matrix
elif viz_type == "Correlation Matrix":
# Get numerical columns
numerical_columns = [col for col in st.session_state.data.columns if
st.session_state.data[col].dtype != 'object']
if len(numerical_columns) > 1:
# Select columns for correlation
selected_columns = st.multiselect(
"Select columns for correlation matrix",
options=numerical_columns,
default=[col for col in numerical_columns if col in [
"toxic", "severe_toxic", "obscene", "threat", "insult", "identity_hate"
]]
)
if selected_columns and len(selected_columns) > 1:
# Calculate correlation
correlation = st.session_state.data[selected_columns].corr()
# Plot correlation matrix
fig, ax = plt.subplots(figsize=(10, 8))
sns.heatmap(
correlation,
annot=True,
cmap='coolwarm',
ax=ax,
cbar=True,
fmt='.2f',
linewidths=0.5
)
plt.title('Correlation Matrix')
st.pyplot(fig)
else:
st.warning("Please select at least two columns for correlation matrix.")
else:
st.warning("Not enough numerical columns for correlation analysis.")
# Add more visualization types as needed
# Prediction visualization
if st.session_state.predictions is not None:
st.subheader("Prediction Visualization")
# Distribution of predictions
st.write("Distribution of Predictions:")
# Plot prediction distribution
fig, ax = plt.subplots(figsize=(10, 6))
if 'toxicity_score' in st.session_state.predictions.columns:
sns.histplot(st.session_state.predictions['toxicity_score'], bins=20, kde=True, ax=ax)
plt.xlabel('Toxicity Score')
plt.ylabel('Frequency')
plt.title('Distribution of Toxicity Scores')
st.pyplot(fig)
# Toxicity threshold analysis
st.write("Toxicity Threshold Analysis:")
threshold = st.slider("Toxicity Threshold", 0.0, 1.0, 0.5, 0.05)
# Calculate metrics at different thresholds
st.session_state.predictions['is_toxic_at_threshold'] = st.session_state.predictions[
'toxicity_score'] > threshold
toxic_at_threshold = st.session_state.predictions['is_toxic_at_threshold'].sum()
total_predictions = len(st.session_state.predictions)
st.write(f"Threshold: {threshold}")
st.write(
f"Toxic comments: {toxic_at_threshold} ({toxic_at_threshold / total_predictions * 100:.2f}%)")
st.write(
f"Non-toxic comments: {total_predictions - toxic_at_threshold} ({(total_predictions - toxic_at_threshold) / total_predictions * 100:.2f}%)")
else:
st.warning("Toxicity scores not found in predictions.")
else:
st.info("Please upload and preprocess data for visualization.")
if __name__ == "__main__":
main() |