app.py

import os
import re
import html
import numpy as np
import torch
import torch.nn as nn
from transformers import BertTokenizerFast, BertForSequenceClassification
import gradio as gr

# ----------------------------------------------------
# Config
# ----------------------------------------------------
MODEL_NAME = "bert-base-uncased"
MAX_LEN = 128

SENTIMENT_MODEL_PATH = "sentiment_bert_best.pt"
SARCASM_MODEL_PATH = "sarcasm_bert_best.pt"

SARCASM_THRESHOLD = 0.6
CORRECTION_CONFIDENCE_THRESHOLD = 0.55

id2sentiment = {0: "Negative", 1: "Neutral", 2: "Positive"}

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
softmax = nn.Softmax(dim=1)

# ----------------------------------------------------
# Minimal cleaning for BERT
# ----------------------------------------------------
def clean_text(text: str) -> str:
    if not isinstance(text, str):
        return ""

    text = html.unescape(text)
    text = re.sub(r"<.*?>", " ", text)
    text = re.sub(r"http\S+|www\.\S+", " ", text)
    text = re.sub(r"\s+", " ", text).strip()

    return text

# ----------------------------------------------------
# Load tokenizer & models
# ----------------------------------------------------
tokenizer = BertTokenizerFast.from_pretrained(MODEL_NAME)

sentiment_model = BertForSequenceClassification.from_pretrained(
    MODEL_NAME, num_labels=3
)
sarcasm_model = BertForSequenceClassification.from_pretrained(
    MODEL_NAME, num_labels=2
)

if os.path.exists(SENTIMENT_MODEL_PATH):
    sentiment_model.load_state_dict(
        torch.load(SENTIMENT_MODEL_PATH, map_location=device)
    )

if os.path.exists(SARCASM_MODEL_PATH):
    sarcasm_model.load_state_dict(
        torch.load(SARCASM_MODEL_PATH, map_location=device)
    )

sentiment_model.to(device).eval()
sarcasm_model.to(device).eval()

# ----------------------------------------------------
# Prediction Helpers
# ----------------------------------------------------
def get_probs(model, text):
    encoding = tokenizer(
        text,
        max_length=MAX_LEN,
        padding="max_length",
        truncation=True,
        return_tensors="pt",
    )

    input_ids = encoding["input_ids"].to(device)
    attention_mask = encoding["attention_mask"].to(device)

    with torch.no_grad():
        outputs = model(input_ids=input_ids, attention_mask=attention_mask)
        probs = softmax(outputs.logits).cpu().numpy()[0]

    return probs

# ----------------------------------------------------
# Fusion Logic (Sarcasm-aware correction)
# ----------------------------------------------------
def fusion_polarity_correction(sarc_probs, sent_probs):
    sarc_pred = int(np.argmax(sarc_probs))
    sarc_conf = float(sarc_probs[sarc_pred])

    sent_pred = int(np.argmax(sent_probs))
    sent_conf = float(sent_probs[sent_pred])

    final_pred = sent_pred
    corrected = False

    if sarc_pred == 1 and sarc_conf >= SARCASM_THRESHOLD:
        if sent_conf >= CORRECTION_CONFIDENCE_THRESHOLD:
            if sent_pred == 2:
                final_pred = 0
                corrected = True
            elif sent_pred == 0:
                final_pred = 2
                corrected = True
            elif sent_pred == 1:
                final_pred = 0
                corrected = True

    return final_pred, corrected, sarc_conf, sent_conf

# ----------------------------------------------------
# Main Prediction Function
# ----------------------------------------------------
def predict_review(text: str):
    if not isinstance(text, str) or not text.strip():
        return "⚠ Please enter valid input text."

    text_clean = clean_text(text)

    sarc_probs = get_probs(sarcasm_model, text_clean)
    sent_probs = get_probs(sentiment_model, text_clean)

    final_pred, corrected, sarc_conf, sent_conf = fusion_polarity_correction(
        sarc_probs, sent_probs
    )

    return {
        "Input": text,
        "Sarcasm Probability": round(float(sarc_probs[1]), 4),
        "Original Sentiment": id2sentiment[int(np.argmax(sent_probs))],
        "Original Confidence": round(sent_conf, 4),
        "Final Sentiment": id2sentiment[final_pred],
        "Correction Applied": corrected,
        "Sentiment Probabilities": {
            "Negative": round(float(sent_probs[0]), 4),
            "Neutral": round(float(sent_probs[1]), 4),
            "Positive": round(float(sent_probs[2]), 4),
        },
    }

# ----------------------------------------------------
# Gradio UI
# ----------------------------------------------------
def gradio_predict(text):
    result = predict_review(text)

    if isinstance(result, str):
        return result

    return (
        f"Input: {result['Input']}\n\n"
        f"Sarcasm Probability: {result['Sarcasm Probability']}\n\n"
        f"Original Sentiment: {result['Original Sentiment']} "
        f"(conf={result['Original Confidence']})\n\n"
        f"Final Sentiment: {result['Final Sentiment']}\n\n"
        f"Correction Applied: {result['Correction Applied']}\n\n"
        f"Sentiment Probabilities:\n"
        f"  Negative: {result['Sentiment Probabilities']['Negative']}\n"
        f"  Neutral:  {result['Sentiment Probabilities']['Neutral']}\n"
        f"  Positive: {result['Sentiment Probabilities']['Positive']}"
    )

with gr.Blocks() as demo:
    gr.Markdown("## 🔍 Sarcasm-Aware Sentiment Analysis (BERT Fusion Model)")

    textbox = gr.Textbox(
        lines=4,
        placeholder="Yeah great, the battery died in 1 hour...",
        label="Enter review text",
    )

    output = gr.Textbox(lines=12, label="Analysis Result")

    btn = gr.Button("Analyze")

    btn.click(gradio_predict, textbox, output)

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