Update app.py

app.py

@@ -1,18 +1,21 @@
 import gradio as gr
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
 import torch
-import easyocr
-from PIL import Image
-import numpy as np
 
-# Hate Speech Model and Labels
-HATE_MODEL_NAME = "cardiffnlp/twitter-roberta-base-hate-multiclass-latest"
-HATE_LABELS = [
+# Sarcasm detection model (public and reliable)
+SARCASM_MODEL_NAME = "j-hartmann/emotion-english-distilroberta-base"
+sarcasm_labels = ["not sarcastic", "sarcastic"]  # simplified mapping
+
+sarcasm_tokenizer = AutoTokenizer.from_pretrained(SARCASM_MODEL_NAME)
+sarcasm_model = AutoModelForSequenceClassification.from_pretrained(SARCASM_MODEL_NAME)
+
+# Hate speech classification DeBERTa model fine-tuned for your labels (hypothetical model)
+HATE_MODEL_NAME = "your-username/deberta-hate-speech-custom"  # replace with your actual fine-tuned model
+hate_labels = [
+    "abusive_words",
+    "harassment",
+    "religious_hate",
     "sexism",
-    "racism",
-    "disability",
-    "sexual_orientation",
-    "religion",
     "other",
     "not_hate"
 ]
@@ -20,49 +23,46 @@ HATE_LABELS = [
 hate_tokenizer = AutoTokenizer.from_pretrained(HATE_MODEL_NAME)
 hate_model = AutoModelForSequenceClassification.from_pretrained(HATE_MODEL_NAME)
 
-reader = easyocr.Reader(['en'], gpu=False)
+def detect_sarcasm(text):
+    inputs = sarcasm_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
+    with torch.no_grad():
+        outputs = sarcasm_model(**inputs)
+        probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
+        pred = torch.argmax(probs).item()
+        confidence = float(probs[0][pred])
+        return sarcasm_labels[pred], confidence
 
-def classify_text(text):
+def classify_hate(text):
     inputs = hate_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
     with torch.no_grad():
         outputs = hate_model(**inputs)
         probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
         pred = torch.argmax(probs).item()
         confidence = float(probs[0][pred])
-        return HATE_LABELS[pred], confidence
+        return hate_labels[pred], confidence
+
+def chatbot(text):
+    if not text or not text.strip():
+        return "Please enter text to analyze."
 
-def ocr_extract(image):
-    if isinstance(image, Image.Image):
-        image = np.array(image)
-    result = reader.readtext(image, detail=0)
-    return ' '.join(result)
+    sarcasm_label, sarcasm_conf = detect_sarcasm(text)
+    if sarcasm_label == "sarcastic":
+        return f"Text is detected as SARCASTIC (Confidence: {sarcasm_conf:.2f}). Hate speech classification is skipped."
 
-def chatbot(image=None, text=None):
-    if image is not None:
-        extracted = ocr_extract(image)
-        if not extracted.strip():
-            return "No text found in image.", None
-        label, confidence = classify_text(extracted)
-        return f"OCR Extracted: {extracted}\nHate Speech: {label} (Confidence: {confidence:.2f})", label
-    elif text and text.strip():
-        label, confidence = classify_text(text)
-        return f"Text: {text}\nHate Speech: {label} (Confidence: {confidence:.2f})", label
-    else:
-        return "Please provide an image or some text.", None
+    hate_label, hate_conf = classify_hate(text)
+    return (
+        f"Text is NOT sarcastic.\n"
+        f"Hate Speech Classification: {hate_label} (Confidence: {hate_conf:.2f})"
+    )
 
 iface = gr.Interface(
     fn=chatbot,
-    inputs=[
-        gr.Image(type="pil", label="Upload Screenshot (optional)"),
-        gr.Textbox(lines=3, placeholder="Or, type/paste text here")
-    ],
-    outputs=[
-        gr.Textbox(label="Prediction"),
-        gr.Label(num_top_classes=len(HATE_LABELS), label="Hate Speech Class"),
-    ],
-    title="Hate Speech Detection Chatbot",
-    description="Detects hate speech categories from text or screenshots."
+    inputs=gr.Textbox(lines=3, placeholder="Enter text"),
+    outputs="text",
+    title="Sarcasm-aware Hate Speech Classifier",
+    description="""First detects sarcasm, and if no sarcasm, classifies hate speech into
+    detailed categories: abusive words, harassment (e.g., body shaming), religious hate, sexism, etc."""
 )
 
 if __name__ == "__main__":
-    iface.launch()
+    iface.launch()