Update app.py

app.py

@@ -5,7 +5,7 @@ from transformers import AutoModelForSequenceClassification, AutoTokenizer
 from transformers import RobertaForSequenceClassification, RobertaTokenizer
 from motif_tagging import detect_motifs
 
-# Load sentiment model
+# custom fine-tuned sentiment model
 sentiment_model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tether-sentiment")
 sentiment_tokenizer = AutoTokenizer.from_pretrained("SamanthaStorm/tether-sentiment")
 
@@ -16,19 +16,15 @@ tokenizer = RobertaTokenizer.from_pretrained(model_name, trust_remote_code=True)
 
 LABELS = [
     "gaslighting", "mockery", "dismissiveness", "control", "guilt_tripping", "apology_baiting", "blame_shifting", "projection",
-    "contradictory_statements", "manipulation", "deflection", "insults", "obscure_formal", "recovery_phase", "non_abusive",
-    "suicidal_threat", "physical_threat", "extreme_control"
+    "contradictory_statements", "manipulation", "deflection", "insults", "obscure_formal", "recovery_phase"
 ]
 
 THRESHOLDS = {
     "gaslighting": 0.25, "mockery": 0.15, "dismissiveness": 0.45, "control": 0.43, "guilt_tripping": 0.15,
     "apology_baiting": 0.2, "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": 2.0, "suicidal_threat": 0.45, "physical_threat": 0.02, "extreme_control": 0.30
+    "manipulation": 0.25, "deflection": 0.30, "insults": 0.34, "obscure_formal": 0.25, "recovery_phase": 0.25
 }
 
-PATTERN_LABELS = LABELS[:15]
-
 EXPLANATIONS = {
     "gaslighting": "Gaslighting involves making someone question their own reality or perceptions...",
     "blame_shifting": "Blame-shifting is when one person redirects the responsibility...",
@@ -40,12 +36,13 @@ 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.",
+    "guilt_tripping": "Guilt-tripping uses guilt to manipulate someone’s actions.",
     "manipulation": "Manipulation deceives to influence or control outcomes.",
     "obscure_formal": "Obscure/formal language manipulates through confusion or superiority."
 }
 
 PATTERN_WEIGHTS = {
-    "gaslighting": 1.3, "control": 1.2, "dismissiveness": 0.8, "non_abusive": 0.0
+    "gaslighting": 1.3, "mockery": 1.2, "control": 1.2, "dismissiveness": 0.8
 }
 
 def custom_sentiment(text):
@@ -57,10 +54,11 @@ def custom_sentiment(text):
     label_map = {0: "supportive", 1: "undermining"}
     return {"label": label_map[label_idx], "score": probs[0][label_idx].item()}
 
-def calculate_abuse_level(scores, thresholds, motif_hits=None):
+def calculate_abuse_level(scores, thresholds, motif_hits=None, flag_multiplier=1.0):
     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
-    return base_score
+    base_score *= flag_multiplier
+    return min(base_score, 100.0)
 
 def interpret_abuse_level(score):
     if score > 80:
@@ -81,24 +79,36 @@ 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(LABELS, scores) 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)
-    top_label = LABELS[np.argmax(scores)]
-    return abuse_level, top_label
+    top_patterns = sorted([(label, score) for label, score in zip(LABELS, scores)], key=lambda x: x[1], reverse=True)[:2]
+    return abuse_level, pattern_labels_used, top_patterns
 
 def analyze_composite(msg1, msg2, msg3, flags):
     thresholds = THRESHOLDS
-    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:
+    messages = [msg1, msg2, msg3]
+    active_messages = [m for m in messages if m.strip()]
+    if not active_messages:
         return "Please enter at least one message."
-    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 = [
+
+    flag_multiplier = 1 + (0.1 * len(flags))  # each checked flag increases weight by 10%
+    results = [analyze_single_message(m, thresholds, flags) for m in active_messages]
+    abuse_scores = [r[0] for r in results]
+    composite_score = round(sum(abuse_scores) / len(abuse_scores), 2)
+    label_sets = [label for result in results for label in result[1]]
+    label_counts = {label: label_sets.count(label) for label in set(label_sets)}
+    top_labels = sorted(label_counts.items(), key=lambda x: x[1], reverse=True)[:2]
+    top_explanations = [EXPLANATIONS.get(label, "") for label, _ in top_labels]
+
+    result = f"These messages show patterns of {', '.join(label for label, _ in top_labels)} and are estimated to be {composite_score}% likely abusive."
+    for expl in top_explanations:
+        if expl:
+            result += f"\n• {expl}"
+    return result
+
+textbox_inputs = [
     gr.Textbox(label="Message 1"),
     gr.Textbox(label="Message 2"),
     gr.Textbox(label="Message 3")
@@ -111,7 +121,7 @@ checkboxes = gr.CheckboxGroup(label="Contextual Flags", choices=[
 
 iface = gr.Interface(
     fn=analyze_composite,
-    inputs=txt_inputs + [checkboxes],
+    inputs=textbox_inputs + [checkboxes],
     outputs=gr.Textbox(label="Results"),
     title="Abuse Pattern Detector (Multi-Message)",
     allow_flagging="manual"