Update app.py

app.py

@@ -15,7 +15,7 @@ def get_emotion_profile(text):
     if isinstance(emotions, list) and isinstance(emotions[0], list):
         emotions = emotions[0]
     return {e['label'].lower(): round(e['score'], 3) for e in emotions}
-
+# Emotion model (no retraining needed)
 emotion_pipeline = hf_pipeline(
     "text-classification",
     model="j-hartmann/emotion-english-distilroberta-base",
@@ -23,7 +23,15 @@ emotion_pipeline = hf_pipeline(
     truncation=True
 )
 
+# --- Timeline Visualization Function ---
 def generate_abuse_score_chart(dates, scores, labels):
+    import matplotlib.pyplot as plt
+    import io
+    from PIL import Image
+    from datetime import datetime
+    import re
+
+    # Determine if all entries are valid dates
     if all(re.match(r"\d{4}-\d{2}-\d{2}", d) for d in dates):
         parsed_x = [datetime.strptime(d, "%Y-%m-%d") for d in dates]
         x_labels = [d.strftime("%Y-%m-%d") for d in parsed_x]
@@ -33,12 +41,13 @@ def generate_abuse_score_chart(dates, scores, labels):
 
     fig, ax = plt.subplots(figsize=(8, 3))
     ax.plot(parsed_x, scores, marker='o', linestyle='-', color='darkred', linewidth=2)
+
     for x, y in zip(parsed_x, scores):
         ax.text(x, y + 2, f"{int(y)}%", ha='center', fontsize=8, color='black')
 
     ax.set_xticks(parsed_x)
     ax.set_xticklabels(x_labels)
-    ax.set_xlabel("")
+    ax.set_xlabel("")  # No axis label
     ax.set_ylabel("Abuse Score (%)")
     ax.set_ylim(0, 105)
     ax.grid(True)
@@ -49,32 +58,305 @@ def generate_abuse_score_chart(dates, scores, labels):
     buf.seek(0)
     return Image.open(buf)
 
+
+# --- Abuse Model ---
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 
 model_name = "SamanthaStorm/tether-multilabel-v3"
 model      = AutoModelForSequenceClassification.from_pretrained(model_name)
 tokenizer  = AutoTokenizer.from_pretrained(model_name, use_fast=False)
 
-LABELS = [...]
-THRESHOLDS = {...}
-PATTERN_WEIGHTS = {...}
-RISK_STAGE_LABELS = {...}
-ESCALATION_QUESTIONS = [...]
-DARVO_PATTERNS = {...}
-DARVO_MOTIFS = [...]
+LABELS = [
+    "blame shifting", "contradictory statements", "control", "dismissiveness",
+    "gaslighting", "guilt tripping", "insults", "obscure language",
+    "projection", "recovery phase", "threat"
+]
+
+THRESHOLDS = {
+    "blame shifting": 0.30, "contradictory statements": 0.30, "control": 0.08, "dismissiveness": 0.12,
+    "gaslighting": 0.09, "guilt tripping": 0.4, "insults": 0.10, "obscure language": 0.55,
+    "projection": 0.09, "recovery phase": 0.20, "threat": 0.15
+}
+
+PATTERN_WEIGHTS = {
+    "gaslighting": 1.5,
+    "control": 1.2,
+    "dismissiveness": 0.7,
+    "blame shifting": 0.8,
+    "guilt tripping": 1.2,
+    "insults": 1.4,
+    "projection": 1.2,
+    "recovery phase": 1.1,
+    "contradictory statements": 0.75,
+    "threat": 1.6  # 🔧 New: raise weight for threat
+}
+RISK_STAGE_LABELS = {
+    1: "🌀 Risk Stage: Tension-Building\nThis message reflects rising emotional pressure or subtle control attempts.",
+    2: "🔥 Risk Stage: Escalation\nThis message includes direct or aggressive patterns, suggesting active harm.",
+    3: "🌧️ Risk Stage: Reconciliation\nThis message reflects a reset attempt—apologies or emotional repair without accountability.",
+    4: "🌸 Risk Stage: Calm / Honeymoon\nThis message appears supportive but may follow prior harm, minimizing it."
+}
+
+ESCALATION_QUESTIONS = [
+    ("Partner has access to firearms or weapons", 4),
+    ("Partner threatened to kill you", 3),
+    ("Partner threatened you with a weapon", 3),
+    ("Partner has ever choked you, even if you considered it consensual at the time", 4),
+    ("Partner injured or threatened your pet(s)", 3),
+    ("Partner has broken your things, punched or kicked walls, or thrown things ", 2),
+    ("Partner forced or coerced you into unwanted sexual acts", 3),
+    ("Partner threatened to take away your children", 2),
+    ("Violence has increased in frequency or severity", 3),
+    ("Partner monitors your calls/GPS/social media", 2)
+]
+DARVO_PATTERNS = {
+    "blame shifting", "projection", "dismissiveness", "guilt tripping", "contradictory statements"
+}
+DARVO_MOTIFS = [
+    "I never said that.", "You’re imagining things.", "That never happened.",
+    "You’re making a big deal out of nothing.", "It was just a joke.", "You’re too sensitive.",
+    "I don’t know what you’re talking about.", "You’re overreacting.", "I didn’t mean it that way.",
+    "You’re twisting my words.", "You’re remembering it wrong.", "You’re always looking for something to complain about.",
+    "You’re just trying to start a fight.", "I was only trying to help.", "You’re making things up.",
+    "You’re blowing this out of proportion.", "You’re being paranoid.", "You’re too emotional.",
+    "You’re always so dramatic.", "You’re just trying to make me look bad.",
+
+    "You’re crazy.", "You’re the one with the problem.", "You’re always so negative.",
+    "You’re just trying to control me.", "You’re the abusive one.", "You’re trying to ruin my life.",
+    "You’re just jealous.", "You’re the one who needs help.", "You’re always playing the victim.",
+    "You’re the one causing all the problems.", "You’re just trying to make me feel guilty.",
+    "You’re the one who can’t let go of the past.", "You’re the one who’s always angry.",
+    "You’re the one who’s always complaining.", "You’re the one who’s always starting arguments.",
+    "You’re the one who’s always making things worse.", "You’re the one who’s always making me feel bad.",
+    "You’re the one who’s always making me look like the bad guy.",
+    "You’re the one who’s always making me feel like a failure.",
+    "You’re the one who’s always making me feel like I’m not good enough.",
+
+    "I can’t believe you’re doing this to me.", "You’re hurting me.",
+    "You’re making me feel like a terrible person.", "You’re always blaming me for everything.",
+    "You’re the one who’s abusive.", "You’re the one who’s controlling.", "You’re the one who’s manipulative.",
+    "You’re the one who’s toxic.", "You’re the one who’s gaslighting me.",
+    "You’re the one who’s always putting me down.", "You’re the one who’s always making me feel bad.",
+    "You’re the one who’s always making me feel like I’m not good enough.",
+    "You’re the one who’s always making me feel like I’m the problem.",
+    "You’re the one who’s always making me feel like I’m the bad guy.",
+    "You’re the one who’s always making me feel like I’m the villain.",
+    "You’re the one who’s always making me feel like I’m the one who needs to change.",
+    "You’re the one who’s always making me feel like I’m the one who’s wrong.",
+    "You’re the one who’s always making me feel like I’m the one who’s crazy.",
+    "You’re the one who’s always making me feel like I’m the one who’s abusive.",
+    "You’re the one who’s always making me feel like I’m the one who’s toxic."
+]
+def get_emotional_tone_tag(emotions, sentiment, patterns, abuse_score):
+    sadness = emotions.get("sadness", 0)
+    joy = emotions.get("joy", 0)
+    neutral = emotions.get("neutral", 0)
+    disgust = emotions.get("disgust", 0)
+    anger = emotions.get("anger", 0)
+    fear = emotions.get("fear", 0)
+
+    # 1. Performative Regret
+    if (
+        sadness > 0.4 and
+        any(p in patterns for p in ["blame shifting", "guilt tripping", "recovery phase"]) and
+        (sentiment == "undermining" or abuse_score > 40)
+    ):
+        return "performative regret"
+
+    # 2. Coercive Warmth
+    if (
+        (joy > 0.3 or sadness > 0.4) and
+        any(p in patterns for p in ["control", "gaslighting"]) and
+        sentiment == "undermining"
+    ):
+        return "coercive warmth"
+
+    # 3. Cold Invalidation
+    if (
+        (neutral + disgust) > 0.5 and
+        any(p in patterns for p in ["dismissiveness", "projection", "obscure language"]) and
+        sentiment == "undermining"
+    ):
+        return "cold invalidation"
+
+    # 4. Genuine Vulnerability
+    if (
+        (sadness + fear) > 0.5 and
+        sentiment == "supportive" and
+        all(p in ["recovery phase"] for p in patterns)
+    ):
+        return "genuine vulnerability"
+
+    # 5. Emotional Threat
+    if (
+        (anger + disgust) > 0.5 and
+        any(p in patterns for p in ["control", "threat", "insults", "dismissiveness"]) and
+        sentiment == "undermining"
+    ):
+        return "emotional threat"
+
+    # 6. Weaponized Sadness
+    if (
+        sadness > 0.6 and
+        any(p in patterns for p in ["guilt tripping", "projection"]) and
+        sentiment == "undermining"
+    ):
+        return "weaponized sadness"
+
+    # 7. Toxic Resignation
+    if (
+        neutral > 0.5 and
+        any(p in patterns for p in ["dismissiveness", "obscure language"]) and
+        sentiment == "undermining"
+    ):
+        return "toxic resignation"
+
+    return None
+def detect_contradiction(message):
+    patterns = [
+        (r"\b(i love you).{0,15}(i hate you|you ruin everything)", re.IGNORECASE),
+        (r"\b(i’m sorry).{0,15}(but you|if you hadn’t)", re.IGNORECASE),
+        (r"\b(i’m trying).{0,15}(you never|why do you)", re.IGNORECASE),
+        (r"\b(do what you want).{0,15}(you’ll regret it|i always give everything)", re.IGNORECASE),
+        (r"\b(i don’t care).{0,15}(you never think of me)", re.IGNORECASE),
+        (r"\b(i guess i’m just).{0,15}(the bad guy|worthless|never enough)", re.IGNORECASE)
+    ]
+    return any(re.search(p, message, flags) for p, flags in patterns)
+    
+def calculate_darvo_score(patterns, sentiment_before, sentiment_after, motifs_found, contradiction_flag=False):
+    # Count all detected DARVO-related patterns
+    pattern_hits = sum(1 for p in patterns if p.lower() in DARVO_PATTERNS)
+
+    # Sentiment delta
+    sentiment_shift_score = max(0.0, sentiment_after - sentiment_before)
+
+    # Match against DARVO motifs more loosely
+    motif_hits = sum(
+        any(phrase.lower() in motif.lower() or motif.lower() in phrase.lower()
+            for phrase in DARVO_MOTIFS)
+        for motif in motifs_found
+    )
+    motif_score = motif_hits / max(len(DARVO_MOTIFS), 1)
+
+    # Contradiction still binary
+    contradiction_score = 1.0 if contradiction_flag else 0.0
+
+    # Final DARVO score
+    return round(min(
+        0.3 * pattern_hits +
+        0.3 * sentiment_shift_score +
+        0.25 * motif_score +
+        0.15 * contradiction_score, 1.0
+    ), 3)
+def detect_weapon_language(text):
+    weapon_keywords = [
+        "knife", "knives", "stab", "cut you", "cutting",
+        "gun", "shoot", "rifle", "firearm", "pistol",
+        "bomb", "blow up", "grenade", "explode",
+        "weapon", "armed", "loaded", "kill you", "take you out"
+    ]
+    text_lower = text.lower()
+    return any(word in text_lower for word in weapon_keywords)
+def get_risk_stage(patterns, sentiment):
+    if "threat" in patterns or "insults" in patterns:
+        return 2
+    elif "recovery phase" in patterns:
+        return 3
+    elif "control" in patterns or "guilt tripping" in patterns:
+        return 1
+    elif sentiment == "supportive" and any(p in patterns for p in ["projection", "dismissiveness"]):
+        return 4
+    return 1
+
+def generate_risk_snippet(abuse_score, top_label, escalation_score, stage):
+    if abuse_score >= 85 or escalation_score >= 16:
+        risk_level = "high"
+    elif abuse_score >= 60 or escalation_score >= 8:
+        risk_level = "moderate"
+    elif stage == 2 and abuse_score >= 40:
+        risk_level = "moderate"  # 🔧 New rule for escalation stage
+    else:
+        risk_level = "low"
+    if isinstance(top_label, str) and " – " in top_label:
+        pattern_label, pattern_score = top_label.split(" – ")
+    else:
+        pattern_label = str(top_label) if top_label is not None else "Unknown"
+        pattern_score = ""
+
+    WHY_FLAGGED = {
+        "control": "This message may reflect efforts to restrict someone’s autonomy, even if it's framed as concern or care.",
+        "gaslighting": "This message could be manipulating someone into questioning their perception or feelings.",
+        "dismissiveness": "This message may include belittling, invalidating, or ignoring the other person’s experience.",
+        "insults": "Direct insults often appear in escalating abusive dynamics and can erode emotional safety.",
+        "threat": "This message includes threatening language, which is a strong predictor of harm.",
+        "blame shifting": "This message may redirect responsibility to avoid accountability, especially during conflict.",
+        "guilt tripping": "This message may induce guilt in order to control or manipulate behavior.",
+        "recovery phase": "This message may be part of a tension-reset cycle, appearing kind but avoiding change.",
+        "projection": "This message may involve attributing the abuser’s own behaviors to the victim.",
+        "default": "This message contains language patterns that may affect safety, clarity, or emotional autonomy."
+    }
 
-# (Leave the rest of your helper functions unchanged)
+    explanation = WHY_FLAGGED.get(pattern_label.lower(), WHY_FLAGGED["default"])
 
+    base = f"\n\n🛑 Risk Level: {risk_level.capitalize()}\n"
+    base += f"This message shows strong indicators of **{pattern_label}**. "
+
+    if risk_level == "high":
+        base += "The language may reflect patterns of emotional control, even when expressed in soft or caring terms.\n"
+    elif risk_level == "moderate":
+        base += "There are signs of emotional pressure or indirect control that may escalate if repeated.\n"
+    else:
+        base += "The message does not strongly indicate abuse, but it's important to monitor for patterns.\n"
+    
+    base += f"\n💡 *Why this might be flagged:*\n{explanation}\n"
+    base += f"\nDetected Pattern: **{pattern_label} ({pattern_score})**\n"
+    base += "🧠 You can review the pattern in context. This tool highlights possible dynamics—not judgments."
+    return base
+def compute_abuse_score(matched_scores, sentiment):
+    if not matched_scores:
+        return 0
+
+    # Weighted average of passed patterns
+    weighted_total = sum(score * weight for _, score, weight in matched_scores)
+    weight_sum = sum(weight for _, _, weight in matched_scores)
+    base_score = (weighted_total / weight_sum) * 100
+
+    # Boost for pattern count
+    pattern_count = len(matched_scores)
+    scale = 1.0 + 0.25 * max(0, pattern_count - 1)  # 1.25x for 2, 1.5x for 3+
+    scaled_score = base_score * scale
+
+    # Pattern floors
+    FLOORS = {
+        "threat": 70,
+        "control": 40,
+        "gaslighting": 30,
+        "insults": 25
+    }
+    floor = max(FLOORS.get(label, 0) for label, _, _ in matched_scores)
+    adjusted_score = max(scaled_score, floor)
+
+    # Sentiment tweak
+    if sentiment == "undermining" and adjusted_score < 50:
+        adjusted_score += 10
+
+    return min(adjusted_score, 100)
+    
+    
 def analyze_single_message(text, thresholds):
     motif_hits, matched_phrases = detect_motifs(text)
+
+    # Get emotion profile
     emotion_profile = get_emotion_profile(text)
     sentiment_score = emotion_profile.get("anger", 0) + emotion_profile.get("disgust", 0)
 
+    # Get model scores first so they can be used in the neutral override
     inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
     with torch.no_grad():
         outputs = model(**inputs)
     scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy()
 
+    # Sentiment override if neutral masks abuse
     if emotion_profile.get("neutral", 0) > 0.85 and any(
         scores[label_idx] > thresholds[LABELS[label_idx]]
         for label_idx in [LABELS.index(l) for l in ["control", "threat", "blame shifting"]]
@@ -83,17 +365,76 @@ def analyze_single_message(text, thresholds):
     else:
         sentiment = "undermining" if sentiment_score > 0.25 else "supportive"
 
+    weapon_flag = detect_weapon_language(text)
+
     adjusted_thresholds = {
         k: v + 0.05 if sentiment == "supportive" else v
         for k, v in thresholds.items()
     }
 
-    passed = {
-        label: score for label, score in zip(LABELS, scores)
+    contradiction_flag = detect_contradiction(text)
+
+    threshold_labels = [
+        label for label, score in zip(LABELS, scores)
         if score > adjusted_thresholds[label]
-    }
+    ]
+
+    motifs = [phrase for _, phrase in matched_phrases]
+
+    darvo_score = calculate_darvo_score(
+        threshold_labels,
+        sentiment_before=0.0,
+        sentiment_after=sentiment_score,
+        motifs_found=motifs,
+        contradiction_flag=contradiction_flag
+    )
 
-    # (Continue unchanged)
+    top_patterns = sorted(
+        [(label, score) for label, score in zip(LABELS, scores)],
+        key=lambda x: x[1],
+        reverse=True
+    )[:2]
+
+    matched_scores = [
+        (label, score, PATTERN_WEIGHTS.get(label, 1.0))
+        for label, score in zip(LABELS, scores)
+        if score > adjusted_thresholds[label]
+    ]
+
+    abuse_score_raw = compute_abuse_score(matched_scores, sentiment)
+    abuse_score = abuse_score_raw
+
+    stage = get_risk_stage(threshold_labels, sentiment) if threshold_labels else 1
+    if weapon_flag and stage < 2:
+        stage = 2
+
+    if weapon_flag:
+        abuse_score_raw = min(abuse_score_raw + 25, 100)
+
+    abuse_score = min(abuse_score_raw, 100 if "threat" in threshold_labels or "control" in threshold_labels else 95)
+
+    # Get tone tag
+    tone_tag = get_emotional_tone_tag(emotion_profile, sentiment, threshold_labels, abuse_score)
+    print(f"Emotional Tone Tag: {tone_tag}")
+
+    # Debug logs
+    print("Emotion Profile:")
+    for emotion, score in emotion_profile.items():
+        print(f"  {emotion.capitalize():10}: {score}")
+    print("\n--- Debug Info ---")
+    print(f"Text: {text}")
+    print(f"Sentiment (via emotion): {sentiment} (score: {round(sentiment_score, 3)})")
+    print("Abuse Pattern Scores:")
+    for label, score in zip(LABELS, scores):
+        passed = "✅" if score > adjusted_thresholds[label] else "❌"
+        print(f"  {label:25} → {score:.3f} {passed}")
+    print(f"Matched for score: {[(l, round(s, 3)) for l, s, _ in matched_scores]}")
+    print(f"Abuse Score Raw: {round(abuse_score_raw, 1)}")
+    print(f"Motifs: {motifs}")
+    print(f"Contradiction: {contradiction_flag}")
+    print("------------------\n")
+
+    return abuse_score, threshold_labels, top_patterns, {"label": sentiment}, stage, darvo_score
 
 def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
     none_selected_checked = answers_and_none[-1]
@@ -118,17 +459,15 @@ def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
         return "Please enter at least one message."
 
     results = [(analyze_single_message(m, THRESHOLDS.copy()), d) for m, d in active]
-
     for result, date in results:
-        assert len(result) == 7, "Unexpected output from analyze_single_message"
-
-    top_labels = [r[0][6] for r in results]
+        assert len(result) == 6, "Unexpected output from analyze_single_message"
+    abuse_scores = [r[0][0] for r in results]
+    top_labels = [r[0][1][0] if r[0][1] else r[0][2][0][0] for r in results]
     top_scores = [r[0][2][0][1] for r in results]
     sentiments = [r[0][3]['label'] for r in results]
     stages = [r[0][4] for r in results]
     darvo_scores = [r[0][5] for r in results]
-    dates_used = [r[1] or "Undated" for r in results]
-    abuse_scores = [r[0][0] for r in results]
+    dates_used = [r[1] or "Undated" for r in results]  # Store dates for future mapping
 
     composite_abuse = int(round(sum(abuse_scores) / len(abuse_scores)))
     top_label = f"{top_labels[0]} – {int(round(top_scores[0] * 100))}%"
@@ -145,6 +484,7 @@ def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
     out = f"Abuse Intensity: {composite_abuse}%\n"
     out += "📊 This reflects the strength and severity of detected abuse patterns in the message(s).\n\n"
 
+# Save this line for later use at the 
     if escalation_score is None:
         escalation_text = "📉 Escalation Potential: Unknown (Checklist not completed)\n"
         escalation_text += "⚠️ *This section was not completed. Escalation potential is unknown.*\n"
@@ -152,20 +492,17 @@ def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
         escalation_text = f"🧨 **Escalation Potential: {risk_level} ({escalation_score}/{sum(w for _, w in ESCALATION_QUESTIONS)})**\n"
         escalation_text += "This score comes directly from the safety checklist and functions as a standalone escalation risk score.\n"
         escalation_text += "It indicates how many serious risk factors are present based on your answers to the safety checklist.\n"
-
     if top_label is None:
         top_label = "Unknown – 0%"
-
     out += generate_risk_snippet(composite_abuse, top_label, escalation_score if escalation_score is not None else 0, most_common_stage)
     out += f"\n\n{stage_text}"
     out += darvo_blurb
-    out += "\n\n" + escalation_text
-
     print(f"DEBUG: avg_darvo = {avg_darvo}")
-    pattern_labels = [r[0][2][0][0] for r in results]
+    pattern_labels = [r[0][2][0][0] for r in results]  # top label for each message
     timeline_image = generate_abuse_score_chart(dates_used, abuse_scores, pattern_labels)
+    out += "\n\n" + escalation_text
     return out, timeline_image
-
+    
 message_date_pairs = [
     (
         gr.Textbox(label=f"Message {i+1}"),