app.py

# =========================
# GRADIO: FastText + BiLSTM + Mixed Rule (HF)
# =========================
import json, pickle, re
import numpy as np
import tensorflow as tf
from tensorflow.keras.preprocessing.sequence import pad_sequences
import gradio as gr

# =========================
# LOAD ARTIFACTS (HF relative paths)
# =========================
MODEL_PATH = "model.keras"
TOK_PATH   = "tokenizer.pkl"
LE_PATH    = "label_encoder.pkl"
CFG_PATH   = "config.json"

model = tf.keras.models.load_model(MODEL_PATH)

with open(TOK_PATH, "rb") as f:
    tokenizer = pickle.load(f)

with open(LE_PATH, "rb") as f:
    le = pickle.load(f)

with open(CFG_PATH, "r", encoding="utf-8") as f:
    cfg = json.load(f)

MAX_LEN = int(cfg["max_len"])

# =========================
# RULE KEYWORDS (STRONG / LIGHT)
# =========================
NEG_LIGHT = re.compile(
    r"(giao(\s*hàng)?|ship|vận\s*chuyển).*(chậm|trễ|lâu|delay)",
    re.I
)

NEG_STRONG = re.compile(
    r"(tệ|kém|lỗi|hỏng|bực|thất\s*vọng|không\s*đáng\s*tiền)",
    re.I
)

POS_STRONG = re.compile(
    r"(rất\s*tốt|rất\s*hài\s*lòng|rất\s*ưng\s*ý|cực\s*kỳ\s*tốt)",
    re.I
)

POS_LIGHT = re.compile(
    r"(tốt|chất\s*lượng|hài\s*lòng|ưng\s*ý|ổn|xịn)",
    re.I
)

# =========================
# PREDICT WITH MIXED LOGIC
# =========================
def predict_label(text: str) -> str:
    text = "" if text is None else str(text).strip()
    if not text:
        return ""

    # model prediction
    seq = tokenizer.texts_to_sequences([text])
    X = pad_sequences(seq, maxlen=MAX_LEN, padding="post", truncating="post")
    prob = model.predict(X, verbose=0)[0]

    prob_map = {le.classes_[i]: float(prob[i]) for i in range(len(prob))}
    p_pos = prob_map.get("Positive", 0.0)
    p_neg = prob_map.get("Negative", 0.0)

    # rule signals
    pos_strong = bool(POS_STRONG.search(text))
    pos_light  = bool(POS_LIGHT.search(text))
    neg_strong = bool(NEG_STRONG.search(text))
    neg_light  = bool(NEG_LIGHT.search(text))

    # decision logic
    if neg_strong and not pos_strong:
        return "Negative"

    if pos_strong and neg_light and not neg_strong:
        return "Positive"

    if pos_light and neg_light:
        return "Neutral"

    return le.classes_[int(np.argmax(prob))]

# =========================
# GRADIO UI
# =========================
demo = gr.Interface(
    fn=predict_label,
    inputs=gr.Textbox(
        lines=3,
        placeholder="Nhập bình luận cần phân tích cảm xúc..."
    ),
    outputs=gr.Textbox(label="Kết quả"),
    title="Sentiment Analysis - FastText + BiLSTM (Tiki Reviews)",
)

if __name__ == "__main__":
    demo.launch()