Update app.py

app.py

@@ -1,16 +1,9 @@
 import gradio as gr
-import re
-from transformers import pipeline as hf_pipeline
+import torch
+from transformers import pipeline as hf_pipeline, AutoModelForSequenceClassification, AutoTokenizer
+from motif_tagging import detect_motifs
 
-# Load SST model
-sst_classifier = hf_pipeline(
-    "text-classification",
-    model="distilbert-base-uncased-finetuned-sst-2-english",
-    top_k=1,
-    truncation=True
-)
-
-# Load emotion classifier
+# ——— 1) Emotion Pipeline ————————————————————————————————————————————————
 emotion_pipeline = hf_pipeline(
     "text-classification",
     model="j-hartmann/emotion-english-distilroberta-base",
@@ -18,98 +11,95 @@ emotion_pipeline = hf_pipeline(
     truncation=True
 )
 
-# Lexicon rules
-negations = {"not", "never", "no", "none", "nobody", "nothing", "neither", "nowhere", "hardly", "scarcely", "barely"}
-amplifiers = {"very", "really", "extremely", "so", "totally", "completely", "absolutely", "utterly", "super"}
-softeners = {"slightly", "somewhat", "a bit", "a little", "mildly", "fairly", "kind of"}
-
-def preprocess_sentiment_text(text):
-    words = text.lower().split()
-    modified = []
-    negate = False
-
-    for word in words:
-        stripped = re.sub(r'\W+', '', word)
-
-        if stripped in negations:
-            negate = True
-            modified.append("<NEG>")
-            continue
-        if stripped in amplifiers:
-            modified.append(f"<AMP>{word}")
-            continue
-        if stripped in softeners:
-            modified.append(f"<SOFT>{word}")
-            continue
-        if negate:
-            modified.append(f"<NEG>{word}")
-            negate = False
-        else:
-            modified.append(word)
-
-    return " ".join(modified)
-
 def get_emotion_profile(text):
-    emotions = emotion_pipeline(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}
-
-def get_emotional_tone_tag(emotions, sentiment, patterns, abuse_score=0):
-    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)
-
-    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"
-    if (joy > 0.3 or sadness > 0.4) and any(p in patterns for p in ["control", "gaslighting"]) and sentiment == "undermining":
-        return "coercive warmth"
-    if (neutral + disgust) > 0.5 and any(p in patterns for p in ["dismissiveness", "projection", "obscure language"]) and sentiment == "undermining":
-        return "cold invalidation"
-    if (sadness + fear) > 0.5 and sentiment == "supportive" and all(p in ["recovery phase"] for p in patterns):
-        return "genuine vulnerability"
-    if (anger + disgust) > 0.5 and any(p in patterns for p in ["control", "threat", "insults", "dismissiveness"]) and sentiment == "undermining":
-        return "emotional threat"
-    if sadness > 0.6 and any(p in patterns for p in ["guilt tripping", "projection"]) and sentiment == "undermining":
-        return "weaponized sadness"
-    if neutral > 0.5 and any(p in patterns for p in ["dismissiveness", "obscure language"]) and sentiment == "undermining":
-        return "toxic resignation"
-
-    return None
-
+    """
+    Returns a dict of { emotion_label: score } for the input text.
+    """
+    results = emotion_pipeline(text)
+    # some pipelines return [[…]]
+    if isinstance(results, list) and isinstance(results[0], list):
+        results = results[0]
+    return {r["label"].lower(): round(r["score"], 3) for r in results}
+
+
+# ——— 2) Abuse-patterns Model ——————————————————————————————————————————————
+model_name = "SamanthaStorm/tether-multilabel-v3"
+model     = AutoModelForSequenceClassification.from_pretrained(model_name)
+tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
+
+LABELS = [
+    "blame shifting", "contradictory statements", "control", "dismissiveness",
+    "gaslighting", "guilt tripping", "insults", "obscure language",
+    "projection", "recovery phase", "threat"
+]
+
+THRESHOLDS = {
+    "blame shifting": 0.28, "contradictory statements": 0.27, "control": 0.08, "dismissiveness": 0.32,
+    "gaslighting":     0.27, "guilt tripping":           0.31, "insults":  0.10, "obscure language": 0.55,
+    "projection":      0.09, "recovery phase":           0.33, "threat":    0.15
+}
+
+
+# ——— 3) Single-message analysis ——————————————————————————————————————————————
 def analyze_message(text):
-    preprocessed = preprocess_sentiment_text(text)
-
-    # Run sentiment classification
-    sst_result = sst_classifier(preprocessed)
-    sentiment_output = sst_result[0] if isinstance(sst_result, list) else sst_result
-    sentiment_label = "supportive" if sentiment_output["label"] == "POSITIVE" else "undermining"
-    sentiment_score = round(sentiment_output["score"] * 100, 2)
-
-    # Run emotion model
-    emotions = get_emotion_profile(text)
-    emotion_summary = "\n".join([f"{k.title()}: {v:.2f}" for k, v in emotions.items()])
+    """
+    Runs motif detection, emotion profiling, and the abuse-pattern classifier.
+    Returns a dict with:
+      - matched_phrases: list of raw regex hits
+      - emotion_profile: { emotion: score }
+      - active_patterns: [ labels above their threshold ]
+    """
+    motif_hits, matched_phrases = detect_motifs(text)
+    emotion_profile = get_emotion_profile(text)
+
+    # get raw model scores
+    toks = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
+    with torch.no_grad():
+        logits = model(**toks).logits.squeeze(0)
+    scores = torch.sigmoid(logits).cpu().numpy()
+
+    # pick up all labels whose score >= threshold
+    active = [lab for lab, sc in zip(LABELS, scores) if sc >= THRESHOLDS[lab]]
+
+    return {
+        "matched_phrases": matched_phrases,
+        "emotion_profile": emotion_profile,
+        "active_patterns": active
+    }
+
+
+# ——— 4) Composite wrapper (for multiple inputs) ———————————————————————————————————
+def analyze_composite(*texts):
+    """
+    Accept multiple text inputs, run analyze_message on each, and
+    return a single formatted string.
+    """
+    outputs = []
+    for idx, txt in enumerate(texts, start=1):
+        if not txt:
+            continue
+        r = analyze_message(txt)
+        block = (
+            f"── Message {idx} ──\n"
+            f"Matched Phrases: {r['matched_phrases']}\n"
+            f"Emotion Profile: {r['emotion_profile']}\n"
+            f"Active Patterns: {r['active_patterns']}\n"
+        )
+        outputs.append(block)
+    return "\n".join(outputs) if outputs else "Please enter at least one message."
 
-    # No abuse patterns yet
-    patterns = []
-    tone_tag = get_emotional_tone_tag(emotions, sentiment_label, patterns)
-    tone_output = tone_tag if tone_tag else "None detected"
 
-    return (
-        f"🧠 Sentiment: {sentiment_label.title()} ({sentiment_score}%)\n\n"
-        f"🎭 Emotional Profile:\n{emotion_summary}\n\n"
-        f"🔍 Tone Tag: {tone_output}"
-    )
+# ——— 5) Gradio interface ————————————————————————————————————————————————
+# adjust how many message inputs you want here:
+message_inputs = [gr.Textbox(label=f"Message {i+1}") for i in range(3)]
 
 iface = gr.Interface(
-    fn=analyze_message,
-    inputs=gr.Textbox(lines=4, placeholder="Paste a message here..."),
-    outputs="text",
-    title="Tether SST + Emotional Tone Tagger",
-    description="Applies lexicon-enhanced preprocessing, classifies sentiment, profiles emotion, and infers tone tags using behavioral logic."
+    fn=analyze_composite,
+    inputs=message_inputs,
+    outputs=gr.Textbox(label="Analysis"),
+    title="Tether Analyzer (no abuse-score / no DARVO)",
+    description="Detects motifs, emotions, and active abuse-patterns only."
 )
 
-iface.launch()
+if __name__ == "__main__":
+    iface.launch()