Spaces:
Sleeping
Sleeping
Update demo with enhanced Hebrew intent model
Browse files- app.py +217 -40
- requirements.txt +4 -3
app.py
CHANGED
|
@@ -1,49 +1,226 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
-
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
| 3 |
import torch
|
| 4 |
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
|
|
|
| 26 |
try:
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
-
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
with gr.Row():
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
|
| 47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
-
demo
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Gradio Demo for Hebrew Intent Classification
|
| 3 |
+
Deploy this as a Hugging Face Space
|
| 4 |
+
"""
|
| 5 |
+
|
| 6 |
import gradio as gr
|
| 7 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
| 8 |
import torch
|
| 9 |
|
| 10 |
+
|
| 11 |
+
class HebrewIntentClassifier:
|
| 12 |
+
def __init__(self):
|
| 13 |
+
# Use your deployed model
|
| 14 |
+
model_name = "Huggingm1r@n/hebrew-intent-classifier"
|
| 15 |
+
|
| 16 |
+
try:
|
| 17 |
+
print("Loading Hebrew Intent Classification model...")
|
| 18 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 19 |
+
self.model = AutoModelForSequenceClassification.from_pretrained(
|
| 20 |
+
model_name)
|
| 21 |
+
self.model.eval()
|
| 22 |
+
print("Model loaded successfully!")
|
| 23 |
+
except Exception as e:
|
| 24 |
+
print(f"Error loading model: {e}")
|
| 25 |
+
raise e
|
| 26 |
+
|
| 27 |
+
def predict(self, text):
|
| 28 |
+
"""Predict intent for Hebrew text"""
|
| 29 |
+
if not text.strip():
|
| 30 |
+
return "Please enter some Hebrew text", {}, "אנא הכנס טקסט בעברית"
|
| 31 |
+
|
| 32 |
try:
|
| 33 |
+
# Tokenize input
|
| 34 |
+
inputs = self.tokenizer(
|
| 35 |
+
text,
|
| 36 |
+
return_tensors="pt",
|
| 37 |
+
padding=True,
|
| 38 |
+
truncation=True,
|
| 39 |
+
max_length=128
|
| 40 |
+
)
|
| 41 |
+
|
| 42 |
+
# Get prediction
|
| 43 |
+
with torch.no_grad():
|
| 44 |
+
outputs = self.model(**inputs)
|
| 45 |
+
logits = outputs.logits
|
| 46 |
+
probabilities = torch.softmax(logits, dim=-1)
|
| 47 |
+
|
| 48 |
+
# Get all predictions
|
| 49 |
+
all_scores = {}
|
| 50 |
+
for i, prob in enumerate(probabilities[0]):
|
| 51 |
+
intent_name = self.model.config.id2label[i]
|
| 52 |
+
all_scores[intent_name] = float(prob)
|
| 53 |
+
|
| 54 |
+
# Get top prediction
|
| 55 |
+
predicted_id = torch.argmax(logits, dim=-1).item()
|
| 56 |
+
predicted_label = self.model.config.id2label[predicted_id]
|
| 57 |
+
confidence = probabilities[0][predicted_id].item()
|
| 58 |
+
|
| 59 |
+
# Format results in Hebrew and English
|
| 60 |
+
intent_translations = {
|
| 61 |
+
"ביטול מנוי": "Cancel Subscription",
|
| 62 |
+
"שאלה כללית": "General Question",
|
| 63 |
+
"שכחת סיסמה": "Password Reset",
|
| 64 |
+
"תמיכה טכנית": "Technical Support"
|
| 65 |
+
}
|
| 66 |
+
|
| 67 |
+
result_text = f"""
|
| 68 |
+
🎯 **כוונה חזויה / Predicted Intent:** {predicted_label}
|
| 69 |
+
🎲 **רמת ביטחון / Confidence:** {confidence:.1%}
|
| 70 |
+
🔤 **תרגום / Translation:** {intent_translations.get(predicted_label, "Unknown")}
|
| 71 |
+
|
| 72 |
+
📊 **כל התחזיות / All Predictions:**
|
| 73 |
+
"""
|
| 74 |
+
|
| 75 |
+
# Sort by confidence and show all
|
| 76 |
+
sorted_scores = sorted(
|
| 77 |
+
all_scores.items(), key=lambda x: x[1], reverse=True)
|
| 78 |
+
for intent, score in sorted_scores:
|
| 79 |
+
bar = "█" * int(score * 20)
|
| 80 |
+
translation = intent_translations.get(intent, "Unknown")
|
| 81 |
+
result_text += f"\\n{intent} ({translation}): {score:.1%} {bar}"
|
| 82 |
+
|
| 83 |
+
return result_text, all_scores, predicted_label
|
| 84 |
+
|
| 85 |
+
except Exception as e:
|
| 86 |
+
return f"Error: {str(e)}", {}, "שגיאה"
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
# Initialize the classifier
|
| 90 |
+
try:
|
| 91 |
+
classifier = HebrewIntentClassifier()
|
| 92 |
+
except Exception as e:
|
| 93 |
+
print(f"Failed to initialize classifier: {e}")
|
| 94 |
+
classifier = None
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def classify_text(text):
|
| 98 |
+
"""Main classification function"""
|
| 99 |
+
if classifier is None:
|
| 100 |
+
return "Model not loaded properly", {}, "Model Error"
|
| 101 |
+
|
| 102 |
+
return classifier.predict(text)
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
def load_example(example_text):
|
| 106 |
+
"""Load example text into the input"""
|
| 107 |
+
return example_text
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
# Create the Gradio interface
|
| 111 |
+
with gr.Blocks(
|
| 112 |
+
title="🇮🇱 Hebrew Intent Classification",
|
| 113 |
+
theme=gr.themes.Soft(),
|
| 114 |
+
css=".rtl { direction: rtl; text-align: right; }"
|
| 115 |
+
) as demo:
|
| 116 |
+
|
| 117 |
+
gr.Markdown("""
|
| 118 |
+
# 🇮🇱 מסווג כוונות עברית / Hebrew Intent Classification
|
| 119 |
+
|
| 120 |
+
## מה זה עושה? / What does this do?
|
| 121 |
+
|
| 122 |
+
מסווג טקסט עברית לפי כוונת הלקוח לאחת מ-4 קטגוריות:
|
| 123 |
+
|
| 124 |
+
**Classifies Hebrew customer service text into 4 categories:**
|
| 125 |
|
| 126 |
+
- 🔐 **שכחת סיסמה** (Password Reset) - בעיות התחברות וסיסמאות
|
| 127 |
+
- ❌ **ביטול מנוי** (Cancel Subscription) - בקשות לביטול שירות
|
| 128 |
+
- ❓ **שאלה כללית** (General Question) - שאלות על מחירים, שירותים, חשבון
|
| 129 |
+
- 🔧 **תמיכה טכנית** (Technical Support) - בעיות טכניות, תקלות, באגים
|
| 130 |
+
""")
|
| 131 |
|
| 132 |
with gr.Row():
|
| 133 |
+
with gr.Column(scale=1):
|
| 134 |
+
gr.Markdown("### 📝 הכנס טקסט עברית / Enter Hebrew Text")
|
| 135 |
+
|
| 136 |
+
text_input = gr.Textbox(
|
| 137 |
+
label="טקסט / Text:",
|
| 138 |
+
placeholder="לדוגמה: שכחתי את הסיסמה שלי",
|
| 139 |
+
lines=4,
|
| 140 |
+
elem_classes=["rtl"],
|
| 141 |
+
info="הכנס טקסט בעברית הקשור לשירות לקוחות"
|
| 142 |
+
)
|
| 143 |
+
|
| 144 |
+
classify_btn = gr.Button(
|
| 145 |
+
"🔍 סווג כוונה / Classify Intent",
|
| 146 |
+
variant="primary",
|
| 147 |
+
size="lg"
|
| 148 |
+
)
|
| 149 |
+
|
| 150 |
+
gr.Markdown("### 📋 דוגמאות לניסוי / Try These Examples:")
|
| 151 |
+
|
| 152 |
+
examples = [
|
| 153 |
+
("שכחתי את הסיסמה שלי", "🔐 שכחת סיסמה"),
|
| 154 |
+
("רוצה לבטל את המנוי", "❌ ביטול מנוי"),
|
| 155 |
+
("כמה עולה החבילה השנתית", "❓ שאלה כללית"),
|
| 156 |
+
("האתר לא עובד לי", "🔧 תמיכה טכנית"),
|
| 157 |
+
("איך אני משנה את האימייל", "❓ שאלה כללית"),
|
| 158 |
+
("יש לי בעיה טכנית באפליקציה", "🔧 תמיכה טכנית"),
|
| 159 |
+
("איך מבטלים את החשבון", "❌ ביטול מנוי"),
|
| 160 |
+
("לא מצליח להיכנס למערכת", "🔐 שכחת סיסמה")
|
| 161 |
+
]
|
| 162 |
|
| 163 |
+
for text, category in examples:
|
| 164 |
+
gr.Button(
|
| 165 |
+
f"{category}: {text}",
|
| 166 |
+
size="sm"
|
| 167 |
+
).click(
|
| 168 |
+
lambda x=text: x,
|
| 169 |
+
outputs=text_input
|
| 170 |
+
)
|
| 171 |
+
|
| 172 |
+
with gr.Column(scale=1):
|
| 173 |
+
gr.Markdown("### 📊 תוצאות / Results")
|
| 174 |
+
|
| 175 |
+
result_output = gr.Markdown(
|
| 176 |
+
value="התוצאות יופיעו כאן / Results will appear here",
|
| 177 |
+
elem_classes=["rtl"]
|
| 178 |
+
)
|
| 179 |
+
|
| 180 |
+
confidence_plot = gr.Label(
|
| 181 |
+
label="ציוני ביטחון / Confidence Scores",
|
| 182 |
+
num_top_classes=4
|
| 183 |
+
)
|
| 184 |
+
|
| 185 |
+
predicted_intent = gr.Textbox(
|
| 186 |
+
label="כוונה חזויה / Predicted Intent",
|
| 187 |
+
interactive=False
|
| 188 |
+
)
|
| 189 |
+
|
| 190 |
+
# Connect the classification function
|
| 191 |
+
classify_btn.click(
|
| 192 |
+
classify_text,
|
| 193 |
+
inputs=[text_input],
|
| 194 |
+
outputs=[result_output, confidence_plot, predicted_intent]
|
| 195 |
+
)
|
| 196 |
+
|
| 197 |
+
# Allow Enter key to trigger classification
|
| 198 |
+
text_input.submit(
|
| 199 |
+
classify_text,
|
| 200 |
+
inputs=[text_input],
|
| 201 |
+
outputs=[result_output, confidence_plot, predicted_intent]
|
| 202 |
+
)
|
| 203 |
+
|
| 204 |
+
gr.Markdown("""
|
| 205 |
+
---
|
| 206 |
+
### 📈 מידע על המודל / Model Information
|
| 207 |
+
|
| 208 |
+
- **בסיס / Base Model:** BERT Multilingual
|
| 209 |
+
- **נתוני אימון / Training Data:** 135 דוגמאות עברית מתחום שירות לקוחות
|
| 210 |
+
- **ביצועים / Performance:** דיוק גבוה >90% על טקסט עברית
|
| 211 |
+
- **קוד מקור / Source:** [GitHub Repository](https://github.com/your-repo)
|
| 212 |
+
|
| 213 |
+
### 🔗 קישורים / Links
|
| 214 |
+
- [Model on Hugging Face](https://huggingface.co/Huggingm1r@n/hebrew-intent-classifier)
|
| 215 |
+
- [Documentation](https://huggingface.co/Huggingm1r@n/hebrew-intent-classifier/blob/main/README.md)
|
| 216 |
+
|
| 217 |
+
Built with ❤️ using Hugging Face Transformers and Gradio
|
| 218 |
+
""")
|
| 219 |
|
| 220 |
+
# Launch the demo
|
| 221 |
+
if __name__ == "__main__":
|
| 222 |
+
demo.launch(
|
| 223 |
+
share=True,
|
| 224 |
+
server_name="0.0.0.0",
|
| 225 |
+
server_port=7860
|
| 226 |
+
)
|
requirements.txt
CHANGED
|
@@ -1,3 +1,4 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
|
|
|
|
|
| 1 |
+
gradio>=4.0.0
|
| 2 |
+
transformers>=4.20.0
|
| 3 |
+
torch>=1.9.0
|
| 4 |
+
numpy
|