Update app.py

app.py

@@ -1,23 +1,48 @@
 import gradio as gr
 from transformers import pipeline
+import re
 
-# Load SST sentiment model (adjust model name if different)
+# Load your SST model pipeline
 sst_classifier = pipeline(
     "text-classification",
-    model="SamanthaStorm/tether-sst",  # replace with your actual SST model name
+    model="SamanthaStorm/tether-sst",  # replace with your model name
     top_k=None,
     truncation=True
 )
 
+# Define lexicons
+negations = {"not", "never", "no", "none", "hardly", "rarely", "barely", "doesn’t", "isn’t", "wasn’t"}
+amplifiers = {"very", "extremely", "really", "absolutely", "so", "incredibly", "deeply", "profoundly"}
+diminishers = {"somewhat", "slightly", "barely", "kind of", "sort of", "a little"}
+
+def apply_lexicon_rules(text):
+    tokens = text.lower().split()
+    lexicon_score = 0
+
+    for i, token in enumerate(tokens):
+        if token in negations:
+            lexicon_score -= 1
+        if token in amplifiers:
+            lexicon_score += 1
+        if token in diminishers:
+            lexicon_score -= 0.5
+
+    return lexicon_score
+
 def classify_sentiment(text):
+    lex_score = apply_lexicon_rules(text)
     result = sst_classifier(text)[0]
     label = result["label"]
-    score = round(result["score"] * 100, 2)
+    base_score = result["score"]
+
+    # Adjust model score based on lexicon features
+    adjusted_score = min(max(base_score + (lex_score * 0.05), 0), 1)
+    score_pct = round(adjusted_score * 100, 2)
 
     if label == "LABEL_0":
-        return f"Supportive ({score}%)"
+        return f"Supportive ({score_pct}%)"
     elif label == "LABEL_1":
-        return f"Undermining ({score}%)"
+        return f"Undermining ({score_pct}%)"
     else:
         return f"Unknown label ({label})"
 
@@ -25,8 +50,8 @@ iface = gr.Interface(
     fn=classify_sentiment,
     inputs=gr.Textbox(lines=4, placeholder="Paste message here..."),
     outputs="text",
-    title="Tether SST Sentiment Analyzer",
-    description="Classifies language as Supportive or Undermining based on tone and intent. Part of the Tether project."
+    title="Tether SST (Lexicon-Enhanced)",
+    description="Combines transformer model with lexicon-based adjustment to classify Supportive or Undermining tone."
 )
 
 iface.launch()