Update app.py

app.py

@@ -4,9 +4,8 @@ import numpy as np
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 from transformers import RobertaForSequenceClassification, RobertaTokenizer
 from motif_tagging import detect_motifs
-from abuse_type_mapping import determine_abuse_type
 
-# custom fine-tuned sentiment model
+# Load sentiment model
 sentiment_model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tether-sentiment")
 sentiment_tokenizer = AutoTokenizer.from_pretrained("SamanthaStorm/tether-sentiment")
 
@@ -29,7 +28,6 @@ THRESHOLDS = {
 }
 
 PATTERN_LABELS = LABELS[:15]
-DANGER_LABELS = LABELS[15:18]
 
 EXPLANATIONS = {
     "gaslighting": "Gaslighting involves making someone question their own reality or perceptions...",
@@ -42,23 +40,12 @@ EXPLANATIONS = {
     "apology_baiting": "Apology-baiting manipulates victims into apologizing for abuser's behavior.",
     "deflection": "Deflection avoids accountability by redirecting blame.",
     "control": "Control restricts autonomy through manipulation or coercion.",
-    "extreme_control": "Extreme control dominates decisions and behaviors entirely.",
-    "physical_threat": "Physical threats signal risk of bodily harm.",
-    "suicidal_threat": "Suicidal threats manipulate others using self-harm threats.",
-    "guilt_tripping": "Guilt-tripping uses guilt to manipulate someone’s actions.",
     "manipulation": "Manipulation deceives to influence or control outcomes.",
-    "non_abusive": "Non-abusive language is respectful and free of coercion.",
     "obscure_formal": "Obscure/formal language manipulates through confusion or superiority."
 }
 
 PATTERN_WEIGHTS = {
-    "physical_threat": 1.5,
-    "suicidal_threat": 1.4,
-    "extreme_control": 1.5,
-    "gaslighting": 1.3,
-    "control": 1.2,
-    "dismissiveness": 0.8,
-    "non_abusive": 0.0
+    "gaslighting": 1.3, "control": 1.2, "dismissiveness": 0.8, "non_abusive": 0.0
 }
 
 def custom_sentiment(text):
@@ -73,9 +60,6 @@ def custom_sentiment(text):
 def calculate_abuse_level(scores, thresholds, motif_hits=None):
     weighted_scores = [score * PATTERN_WEIGHTS.get(label, 1.0) for label, score in zip(LABELS, scores) if score > thresholds[label]]
     base_score = round(np.mean(weighted_scores) * 100, 2) if weighted_scores else 0.0
-    motif_hits = motif_hits or []
-    if any(label in motif_hits for label in {"physical_threat", "suicidal_threat", "extreme_control"}):
-        base_score = max(base_score, 75.0)
     return base_score
 
 def interpret_abuse_level(score):
@@ -97,28 +81,22 @@ def analyze_single_message(text, thresholds, motif_flags):
     with torch.no_grad():
         outputs = model(**inputs)
     scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy()
-    threshold_labels = [label for label, score in zip(PATTERN_LABELS, scores[:15]) if score > adjusted_thresholds[label]]
-    phrase_labels = [label for label, _ in matched_phrases]
-    pattern_labels_used = list(set(threshold_labels + phrase_labels))
     abuse_level = calculate_abuse_level(scores, adjusted_thresholds, motif_hits)
-    abuse_description = interpret_abuse_level(abuse_level)
-    top_patterns = sorted([(label, score) for label, score in zip(PATTERN_LABELS, scores[:15]) if label != "non_abusive"], key=lambda x: x[1], reverse=True)[:2]
-    pattern_expl = "\n".join([f"• {label.replace('_', ' ').title()}: {EXPLANATIONS.get(label)}" for label, _ in top_patterns])
-    return abuse_level, abuse_description, pattern_expl
+    top_label = LABELS[np.argmax(scores)]
+    return abuse_level, top_label
 
 def analyze_composite(msg1, msg2, msg3, flags):
     thresholds = THRESHOLDS
-    results = [analyze_single_message(m, thresholds, flags) for m in [msg1, msg2, msg3] if m.strip()]
+    messages = [m for m in [msg1, msg2, msg3] if m.strip()]
+    results = [analyze_single_message(m, thresholds, flags) for m in messages]
     if not results:
         return "Please enter at least one message."
-    result_lines = []
-    total_score = 0
-    for i, (score, desc, patterns) in enumerate(results, 1):
-        total_score += score
-        result_lines.append(f"Message {i}: {score:.2f}% – {desc}\n{patterns}\n")
-    composite = round(total_score / len(results), 2)
-    result_lines.append(f"\nComposite Abuse Score: {composite}%")
-    return "\n\n".join(result_lines)
+    abuse_scores = [score for score, _ in results]
+    labels = [label.replace("_", " ") for _, label in results]
+    composite = round(np.mean(abuse_scores), 2)
+    abuse_desc = interpret_abuse_level(composite)
+    top_labels = ", ".join(labels)
+    return f"The messages analyzed show signs of {top_labels}, indicating a {composite}% likelihood of abusive communication ({abuse_desc})."
 
 txt_inputs = [
     gr.Textbox(label="Message 1"),
@@ -140,4 +118,4 @@ iface = gr.Interface(
 )
 
 if __name__ == "__main__":
-    iface.launch()
+    iface.launch()