Update app.py

app.py

@@ -1,8 +1,10 @@
 # ================================
 # 🪞 MoodMirror+ — Conversational Emotional Self-Care
-# Dataset-only: trains a TF-IDF + OneVsRest Logistic Regression on GoEmotions
-# Cache du modèle + DB dans /data quand dispo (HF Spaces: activer Persistent storage)
-# Toujours donner au moins 1 conseil + parfois une citation
+# Dataset-only: TF-IDF + OneVsRest Logistic Regression on GoEmotions
+# - Persists model & SQLite DB to /data (enable Persistent storage on HF Spaces)
+# - Always give at least one advice tip; sometimes add a quote
+# - Implicit emotion recognition (emoji/slang/negations hints)
+# - Dynamic threshold + Top-1 fallback
 # ================================
 import os
 import re
@@ -30,7 +32,7 @@ DATA_DIR = os.getenv("MM_DATA_DIR", _pick_data_dir())
 os.makedirs(DATA_DIR, exist_ok=True)
 DB_PATH = os.path.join(DATA_DIR, "moodmirror.db")
 MODEL_PATH = os.path.join(DATA_DIR, "goemo_sklearn.joblib")
-MODEL_VERSION = "v1-tfidf-lr-ovr"
+MODEL_VERSION = "v2-tfidf-lr-ovr-implicit"  # bump when training recipe changes
 
 print(f"[MM] Using data dir: {DATA_DIR}")
 print(f"[MM] SQLite path:   {DB_PATH}")
@@ -238,7 +240,77 @@ GOEMO_TO_APP = {
     "sadness": "sadness", "surprise": "neutral", "neutral": "neutral",
 }
 
-THRESHOLD = 0.30
+# --- Threshold & implicit-emotion controls ---
+THRESHOLD = 0.30          # standard selection threshold
+MIN_THRESHOLD = 0.12       # floor if model is unsure
+TOP1_FALLBACK = True       # ensure at least one label if nothing passes threshold
+
+# --- Implicit cues (emojis, slang, negations, intensity) ---
+EMOJI_HINTS = {
+    "😢": "sadness", "😭": "sadness", "😞": "sadness", "💔": "grief",
+    "😡": "anger", "🤬": "anger", "😤": "anger",
+    "😱": "fear", "😨": "fear", "😰": "fear",
+    "😌": "relief", "🙂": "joy", "😊": "joy", "😄": "joy", "😍": "love",
+    "🤗": "love", "💖": "love", "🙏": "gratitude",
+    "💤": "boredom", "🥱": "boredom", "🤔": "curiosity",
+    "😐": "neutral", "😶": "neutral"
+}
+SLANG_HINTS = {
+    "idk": "confusion", "meh": "boredom", "ugh": "annoyance", "nah": "disapproval",
+    "wtf": "anger", "omg": "surprise", "lol": "amusement", "lmao": "amusement",
+    "miss you": "grief", "miss her": "grief", "miss him": "grief",
+    "im fine": "sadness", "i'm fine": "sadness"
+}
+NEGATION_PATTERNS = [
+    ("not happy", "sadness"),
+    ("not okay", "sadness"),
+    ("not ok", "sadness"),
+    ("not fine", "sadness"),
+    ("no hope", "sadness"),
+    ("no energy", "sadness"),
+    ("no motivation", "boredom"),
+    ("not safe", "fear"),
+    ("not calm", "nervousness"),
+    ("not sure", "confusion"),
+]
+INTENSIFIERS = ["!!", "!!!", "?!", "?!?", "soooo", "very", "really", "super", "extremely"]
+
+def _lower(s: str) -> str:
+    return s.lower() if isinstance(s, str) else ""
+
+def augment_text_for_classifier(text: str) -> str:
+    """
+    Inject 'hint tokens' based on implicit cues (emojis/slang/negations/intensity)
+    to help TF-IDF pick up emotions not explicitly named.
+    """
+    if not text:
+        return text
+    t = _lower(text)
+    hints = []
+
+    # emojis
+    for ch in text:
+        if ch in EMOJI_HINTS:
+            hints.append(EMOJI_HINTS[ch])
+
+    # slang
+    for k, v in SLANG_HINTS.items():
+        if k in t:
+            hints.append(v)
+
+    # common negations
+    for pat, lab in NEGATION_PATTERNS:
+        if pat in t:
+            hints.append(lab)
+
+    # intensity: duplicate hints to weigh more
+    if hints and any(x in t for x in INTENSIFIERS):
+        hints = hints + hints
+
+    if hints:
+        hint_tokens = " ".join([f"emo_{h}" for h in hints])
+        return text + " " + hint_tokens
+    return text
 
 # ---------------- SQLite helpers ----------------
 def get_conn():
@@ -317,15 +389,36 @@ except Exception as e:
 
 # ---------------- Emotion detection ----------------
 def classify_text(text: str):
-    if not CLASSIFIER:
+    """
+    Returns [(label_name, prob), ...] sorted desc.
+    - Augments input with implicit-cue tokens
+    - Uses dynamic threshold with a safety floor
+    - Top-1 fallback ensures at least one label
+    """
+    if not CLASSIFIER or not LABEL_NAMES:
         return []
+
+    augmented = augment_text_for_classifier(text)
+
     try:
-        proba = CLASSIFIER.predict_proba([text])[0]
+        proba = CLASSIFIER.predict_proba([augmented])[0]
     except AttributeError:
         from scipy.special import expit
-        proba = expit(CLASSIFIER.decision_function([text])[0])
-    idxs = [i for i, p in enumerate(proba) if p >= THRESHOLD]
+        scores = CLASSIFIER.decision_function([augmented])[0]
+        proba = expit(scores)
+
+    # Dynamic threshold
+    maxp = float(max(proba)) if len(proba) else 0.0
+    thr = THRESHOLD if maxp >= THRESHOLD else max(MIN_THRESHOLD, maxp * 0.8)
+
+    idxs = [i for i, p in enumerate(proba) if p >= thr]
     idxs.sort(key=lambda i: proba[i], reverse=True)
+
+    # Top-1 fallback
+    if not idxs and TOP1_FALLBACK and len(proba):
+        top1 = int(max(range(len(proba)), key=lambda i: proba[i]))
+        idxs = [top1]
+
     return [(LABEL_NAMES[i], float(proba[i])) for i in idxs]
 
 def detect_emotions(text: str):
@@ -336,7 +429,7 @@ def detect_emotions(text: str):
     for label, p in chosen:
         app = GOEMO_TO_APP.get(label.lower(), "neutral")
         bucket[app] = max(bucket.get(app, 0.0), p)
-    return max(bucket, key=bucket.get)
+    return max(bucket, key=bucket.get) if bucket else "neutral"
 
 # ---------------- Reply composer ----------------
 def compose_support(main_emotion: str, is_first_msg: bool) -> str: