Update app.py

app.py

@@ -1,62 +1,70 @@
 import gradio as gr
 import torch
-from transformers import RobertaForSequenceClassification, RobertaTokenizer
+from transformers import RobertaForSequenceClassification, RobertaTokenizer, pipeline
 import numpy as np
-from transformers import pipeline
 
-# Load sentiment analysis model
+# Load sentiment model
 sentiment_analyzer = pipeline("sentiment-analysis")
 
-# Load model and tokenizer
+# Load abuse pattern model
 model_name = "SamanthaStorm/abuse-pattern-detector-v2"
 model = RobertaForSequenceClassification.from_pretrained(model_name, trust_remote_code=True)
 tokenizer = RobertaTokenizer.from_pretrained(model_name, trust_remote_code=True)
 
+# Labels
 LABELS = [
     "gaslighting", "mockery", "dismissiveness", "control",
     "guilt_tripping", "apology_baiting", "blame_shifting", "projection",
     "contradictory_statements", "manipulation", "deflection", "insults",
-    "obscure_formal", "recovery_phase", "non_abusive",
+    "obscure_formal", "recovery_phase", "non_abusive", 
     "suicidal_threat", "physical_threat", "extreme_control"
 ]
+PATTERN_LABELS = LABELS[:15]
+DANGER_LABELS = LABELS[15:]
 
+# Thresholds
 THRESHOLDS = {
-    "gaslighting": 0.25, "mockery": 0.15, "dismissiveness": 0.30,
-    "control": 0.43, "guilt_tripping": 0.19, "apology_baiting": 0.45,
-    "blame_shifting": 0.23, "projection": 0.50, "contradictory_statements": 0.25,
-    "manipulation": 0.25, "deflection": 0.30, "insults": 0.34,
-    "obscure_formal": 0.25, "recovery_phase": 0.25, "non_abusive": 0.70,
-    "suicidal_threat": 0.45, "physical_threat": 0.20, "extreme_control": 0.36
+    "gaslighting": 0.25,
+    "mockery": 0.15,
+    "dismissiveness": 0.30,
+    "control": 0.43,
+    "guilt_tripping": 0.19,
+    "apology_baiting": 0.45,
+    "blame_shifting": 0.23,
+    "projection": 0.50,
+    "contradictory_statements": 0.25,
+    "manipulation": 0.25,
+    "deflection": 0.30,
+    "insults": 0.34,
+    "obscure_formal": 0.25,
+    "recovery_phase": 0.25,
+    "non_abusive": 0.70,
+    "suicidal_threat": 0.45,
+    "physical_threat": 0.20,
+    "extreme_control": 0.36
 }
 
-PATTERN_LABELS = LABELS[:15]
-DANGER_LABELS = LABELS[15:]
-
+# Explanations
 EXPLANATIONS = {
-    "gaslighting": "Gaslighting involves making someone question their own reality or perceptions.",
-    "blame_shifting": "Blame-shifting is when one person redirects responsibility onto someone else.",
-    "projection": "Projection accuses the victim of behaviors the abuser exhibits themselves.",
-    "dismissiveness": "Dismissiveness belittles or ignores another person’s thoughts or feelings.",
-    "mockery": "Mockery involves ridicule or sarcasm meant to humiliate.",
-    "recovery_phase": "Recovery phase invalidates someone’s healing process or needs.",
-    "insults": "Insults are derogatory remarks meant to degrade or attack.",
-    "apology_baiting": "Apology-baiting manipulates someone into apologizing for the abuser’s actions.",
-    "deflection": "Deflection shifts responsibility or changes the subject to avoid blame.",
-    "control": "Control includes behavior that limits another’s autonomy or freedom.",
-    "extreme_control": "Extreme control is highly manipulative dominance over another’s choices.",
-    "physical_threat": "Physical threats suggest or state a risk of bodily harm.",
-    "suicidal_threat": "Suicidal threats use self-harm as a way to manipulate others.",
-    "guilt_tripping": "Guilt-tripping makes someone feel guilty for things they didn’t cause.",
-    "manipulation": "Manipulation influences someone’s behavior through deceptive emotional tactics.",
-    "non_abusive": "Non-abusive language is respectful, supportive, and healthy."
+    "gaslighting": "Gaslighting involves making someone question their own reality or perceptions, often causing them to feel confused or insecure.",
+    "blame_shifting": "Blame-shifting is when one person redirects the responsibility for an issue onto someone else, avoiding accountability.",
+    "projection": "Projection involves accusing the victim of behaviors or characteristics that the abuser themselves exhibit.",
+    "dismissiveness": "Dismissiveness is the act of belittling or disregarding another person's thoughts, feelings, or experiences.",
+    "mockery": "Mockery involves ridiculing or making fun of someone in a hurtful way, often with the intent to humiliate them.",
+    "recovery_phase": "Recovery phase refers to dismissing or invalidating someone’s process of emotional healing, or ignoring their need for support.",
+    "insults": "Insults are derogatory remarks aimed at degrading or humiliating someone, often targeting their personal traits or character.",
+    "apology_baiting": "Apology-baiting is when the abuser manipulates the victim into apologizing for something the abuser caused or did wrong.",
+    "deflection": "Deflection is the act of avoiding responsibility or shifting focus away from one's own actions, often to avoid accountability.",
+    "control": "Control tactics are behaviors that restrict or limit someone's autonomy, often involving domination, manipulation, or coercion.",
+    "extreme_control": "Extreme control involves excessive manipulation or domination over someone’s actions, decisions, or behaviors.",
+    "physical_threat": "Physical threats involve any indication or direct mention of harm to someone’s physical well-being, often used to intimidate or control.",
+    "suicidal_threat": "Suicidal threats are statements made to manipulate or control someone by making them feel responsible for the abuser’s well-being.",
+    "guilt_tripping": "Guilt-tripping involves making someone feel guilty or responsible for things they didn’t do, often to manipulate their behavior.",
+    "manipulation": "Manipulation refers to using deceptive tactics to control or influence someone’s emotions, decisions, or behavior to serve the manipulator’s own interests.",
+    "non_abusive": "Non-abusive language is communication that is respectful, empathetic, and free of harmful behaviors or manipulation."
 }
 
-def calculate_abuse_level(scores, thresholds):
-    triggered_scores = [score for label, score in zip(LABELS, scores) if score > thresholds[label]]
-    if not triggered_scores:
-        return 0.0
-    return round(np.mean(triggered_scores) * 100, 2)
-
+# Abuse level interpretation
 def interpret_abuse_level(score):
     if score > 80:
         return "Extreme / High Risk"
@@ -69,6 +77,7 @@ def interpret_abuse_level(score):
     else:
         return "Very Low / Likely Safe"
 
+# Main analysis
 def analyze_messages(input_text, context_flags):
     input_text = input_text.strip()
     if not input_text:
@@ -79,78 +88,75 @@ def analyze_messages(input_text, context_flags):
     sentiment_label = sentiment['label']
     sentiment_score = sentiment['score']
 
-    # Threshold adjustment for negative tone
+    # Adjust thresholds if negative tone
     adjusted_thresholds = THRESHOLDS.copy()
-    if sentiment_label == "NEGATIVE":
-        adjusted_thresholds = {key: val * 0.8 for key, val in THRESHOLDS.items()}
+    if sentiment_label.upper() == "NEGATIVE":
+        adjusted_thresholds = {k: v * 0.8 for k, v in THRESHOLDS.items()}
 
-    # Tokenization and prediction
+    # Run model
     inputs = tokenizer(input_text, return_tensors="pt", truncation=True, padding=True)
     with torch.no_grad():
         outputs = model(**inputs)
     scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy()
 
-    # Pattern and danger flags from model
+    # Pattern & danger from model
     pattern_count = sum(score > adjusted_thresholds[label] for label, score in zip(PATTERN_LABELS, scores[:15]))
     danger_flag_count = sum(score > adjusted_thresholds[label] for label, score in zip(DANGER_LABELS, scores[15:]))
 
-    # Add contextual danger from checkboxes
+    # Add checkbox context flags
     if context_flags and len(context_flags) >= 2:
         danger_flag_count += 1
 
-    # Non-abusive override
+    # Override if non-abusive
     non_abusive_score = scores[LABELS.index('non_abusive')]
     if non_abusive_score > adjusted_thresholds['non_abusive']:
         return "This message is classified as non-abusive."
 
-    # Abuse level
-    abuse_level = calculate_abuse_level(scores, THRESHOLDS)
+    # Abuse score
+    triggered_scores = [score for label, score in zip(LABELS, scores) if score > adjusted_thresholds[label]]
+    abuse_level = round(np.mean(triggered_scores) * 100, 2) if triggered_scores else 0.0
     abuse_description = interpret_abuse_level(abuse_level)
 
+    # Top patterns
+    scored_patterns = [(label, score) for label, score in zip(PATTERN_LABELS, scores[:15])]
+    top_patterns = sorted(scored_patterns, key=lambda x: x[1], reverse=True)[:2]
+    top_pattern_explanations = "\n".join(
+        [f"• {label.replace('_', ' ').title()}: {EXPLANATIONS.get(label, 'No explanation available.')}" for label, _ in top_patterns]
+    )
+
     # Resources
     if danger_flag_count >= 2:
-        resources = "⚠️ Your responses indicate elevated danger. Please consider seeking support immediately through a domestic violence hotline or trusted professional."
+        resources = "Immediate assistance recommended. Please seek professional help or contact emergency services."
     else:
         resources = "For more information on abuse patterns, consider reaching out to support groups or professional counselors."
 
-    # Top patterns with definitions
-    scored_patterns = [(label, score) for label, score in zip(PATTERN_LABELS, scores[:15])]
-    top_patterns = sorted(scored_patterns, key=lambda x: x[1], reverse=True)[:2]
-    top_pattern_explanations = "\n".join([
-        f"• {label.replace('_', ' ').title()}: {EXPLANATIONS.get(label, 'No explanation available.')}"
-        for label, _ in top_patterns
-    ])
-
+    # Result
     result = (
         f"Abuse Risk Score: {abuse_level}% – {abuse_description}\n\n"
         f"Most Likely Patterns:\n{top_pattern_explanations}\n\n"
         f"⚠️ Critical Danger Flags Detected: {danger_flag_count} of 3\n"
         "The Danger Assessment is a validated tool that helps identify serious risk in intimate partner violence. "
-        "It flags communication patterns associated with increased risk of severe harm.\n\n"
-        f"Resources: {resources}\n\n"
+        "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} \n\n"
         f"Sentiment: {sentiment_label} (Confidence: {sentiment_score*100:.2f}%)"
     )
-
     return result
 
-# Launch interface
+# Interface
 iface = gr.Interface(
     fn=analyze_messages,
     inputs=[
-        gr.Textbox(lines=10, placeholder="Enter message here..."),
+        gr.Textbox(lines=10, placeholder="Enter message here...", label="input_text"),
         gr.CheckboxGroup(
+            ["They've threatened harm", "They isolate me", "I've changed my behavior out of fear",
+             "They monitor/follow me", "I feel unsafe when alone with them"],
             label="Do any of these apply to your situation?",
-            choices=[
-                "They’ve threatened harm",
-                "They isolate me",
-                "I’ve changed my behavior out of fear",
-                "They monitor/follow me",
-                "I feel unsafe when alone with them"
-            ]
+            type="value"
         )
     ],
     outputs=gr.Textbox(label="Analysis Result"),
-    title="Abuse Pattern Detector",
+    title="Abuse Pattern Detector"
 )
 
 if __name__ == "__main__":