Update app.py

app.py

@@ -2,10 +2,10 @@ import gradio as gr
 import re
 from transformers import pipeline as hf_pipeline
 
-# Load SST sentiment classifier
+# Load SST model (temporary baseline)
 sst_classifier = hf_pipeline(
     "text-classification",
-    model="SamanthaStorm/tether-sst",
+    model="distilbert-base-uncased-finetuned-sst-2-english",
     top_k=None,
     truncation=True
 )
@@ -18,29 +18,45 @@ emotion_pipeline = hf_pipeline(
     truncation=True
 )
 
-# Load Tether's abuse pattern classifier
-abuse_classifier = hf_pipeline(
-    "text-classification",
-    model="SamanthaStorm/tether-multilabel-v2",
-    top_k=None,
-    truncation=True
-)
+# Lexicon enhancement preprocessing
+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"}
 
-# Extract abuse pattern labels
-def get_abuse_patterns(text, threshold=0.5):
-    results = abuse_classifier(text)
-    if isinstance(results, list) and isinstance(results[0], list):
-        results = results[0]
-    return [r['label'] for r in results if r['score'] >= threshold]
+def preprocess_sentiment_text(text):
+    words = text.lower().split()
+    modified = []
+    negate = False
 
-# Emotion extraction
+    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)
+
+# Emotion mapping
 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}
 
-# Emotional tone tagging based on behavior + emotion + sentiment
+# Tone tagging logic
 def get_emotional_tone_tag(emotions, sentiment, patterns, abuse_score=0):
     sadness = emotions.get("sadness", 0)
     joy = emotions.get("joy", 0)
@@ -49,67 +65,35 @@ def get_emotional_tone_tag(emotions, sentiment, patterns, abuse_score=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)
-    ):
+    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"
-    ):
+    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"
-    ):
+    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)
-    ):
+    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"
-    ):
+    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"
-    ):
+    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"
-    ):
+    if neutral > 0.5 and any(p in patterns for p in ["dismissiveness", "obscure language"]) and sentiment == "undermining":
         return "toxic resignation"
 
     return None
 
-# Analysis logic
+# Main function
 def analyze_message(text):
-    # Sentiment
-    sst_result = sst_classifier(text)[0]
-    sentiment_label = "supportive" if sst_result["label"] == "LABEL_0" else "undermining"
-    sentiment_score = round(sst_result["score"] * 100, 2)
+    preprocessed = preprocess_sentiment_text(text)
+    sentiment_result = sst_classifier(preprocessed)[0]
+    sentiment_label = "supportive" if sentiment_result["label"] == "POSITIVE" else "undermining"
+    sentiment_score = round(sentiment_result["score"] * 100, 2)
 
-    # Emotions
     emotions = get_emotion_profile(text)
     emotion_summary = "\n".join([f"{k.title()}: {v:.2f}" for k, v in emotions.items()])
 
-    # Abuse patterns (used internally)
-    patterns = get_abuse_patterns(text)
-
-    # Tone tag
-    tone_tag = get_emotional_tone_tag(emotions, sentiment_label, patterns)
+    # Temporarily pass empty abuse pattern list until Tether model is added
+    tone_tag = get_emotional_tone_tag(emotions, sentiment_label, patterns=[])
     tone_output = tone_tag if tone_tag else "None detected"
 
     return (
@@ -118,13 +102,13 @@ def analyze_message(text):
         f"🔍 Tone Tag: {tone_output}"
     )
 
-# Gradio UI
+# Interface
 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="Detects Supportive vs Undermining sentiment, emotion profile, and custom tone tags using behavioral logic."
+    description="Applies lexicon-enhanced preprocessing, classifies sentiment, profiles emotion, and infers tone tags based on behavior logic."
 )
 
 iface.launch()