Update app.py

app.py

@@ -2,6 +2,10 @@ import gradio as gr
 import torch
 from transformers import RobertaForSequenceClassification, RobertaTokenizer
 import numpy as np
+from transformers import pipeline
+
+# Load sentiment analysis model
+sentiment_analyzer = pipeline("sentiment-analysis")
 
 # Load model and tokenizer with trust_remote_code in case it's needed
 model_name = "SamanthaStorm/abuse-pattern-detector-v2"
@@ -83,7 +87,18 @@ def analyze_messages(input_text):
     input_text = input_text.strip()
     if not input_text:
         return "Please enter a message for analysis.", None
-
+        
+     # Sentiment analysis
+    sentiment = sentiment_analyzer(input_text)[0]  # Sentiment result
+    sentiment_label = sentiment['label']
+    sentiment_score = sentiment['score']
+    
+     # Adjust thresholds based on sentiment
+    adjusted_thresholds = THRESHOLDS.copy()
+    if sentiment_label == "NEGATIVE":
+        # Lower thresholds for negative sentiment
+        adjusted_thresholds = {key: val * 0.8 for key, val in THRESHOLDS.items()}  # Example adjustment
+        
     # Tokenize input and generate model predictions
     inputs = tokenizer(input_text, return_tensors="pt", truncation=True, padding=True)
     with torch.no_grad():
@@ -91,8 +106,14 @@ def analyze_messages(input_text):
     scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy()
 
     # Count the number of triggered abuse pattern and danger flags based on thresholds
-    pattern_count = sum(score > THRESHOLDS[label] for label, score in zip(PATTERN_LABELS, scores[:14]))
-    danger_flag_count = sum(score > THRESHOLDS[label] for label, score in zip(DANGER_LABELS, scores[14:17]))
+    pattern_count = sum(score > adjusted_thresholds[label] for label, score in zip(PATTERN_LABELS, scores[:14]))
+    danger_flag_count = sum(score > adjusted_thresholds[label] for label, score in zip(DANGER_LABELS, scores[14:17]))
+
+    # Check if 'non_abusive' label is triggered
+    non_abusive_score = scores[LABELS.index('non_abusive')]
+    if non_abusive_score > adjusted_thresholds['non_abusive']:
+        # If non-abusive threshold is met, return a non-abusive classification
+        return "This message is classified as non-abusive."
 
     # Build formatted raw score display
     score_lines = [
@@ -127,6 +148,7 @@ def analyze_messages(input_text):
         "It flags communication patterns associated with increased risk of severe harm. "
         "For more info, consider reaching out to support groups or professionals.\n\n"
         f"Resources: {resources}"
+        f"Sentiment: {sentiment_label} (Confidence: {sentiment_score*100:.2f}%)"
     )
 
     # Return both a text summary and a JSON-like dict of scores per label