Upload classifier.py with huggingface_hub
Browse files- classifier.py +365 -0
classifier.py
ADDED
@@ -0,0 +1,365 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# -*- coding: utf-8 -*-
|
3 |
+
|
4 |
+
# Path to the locally fine-tuned model
|
5 |
+
LOCAL_MODEL_PATH = "./models/finetuned_classification"
|
6 |
+
|
7 |
+
# Hugging Face model name (fallback)
|
8 |
+
MODEL_NAME = "rmtariq/malay_classification"
|
9 |
+
|
10 |
+
# Categories from the new dataset
|
11 |
+
CATEGORIES = ["Politik", "Perpaduan", "Keluarga", "Belia", "Perumahan", "Internet", "Pengguna", "Makanan", "Pekerjaan", "Pengangkutan", "Sukan", "Ekonomi", "Hiburan", "Jenayah", "Alam Sekitar", "Teknologi", "Pendidikan", "Agama", "Sosial", "Kesihatan", "Halal"]
|
12 |
+
|
13 |
+
"""
|
14 |
+
Claim Classifier
|
15 |
+
---------------
|
16 |
+
|
17 |
+
Classifies claims based on priority index data, sentiment analysis, and content patterns.
|
18 |
+
Also provides functions for classifying claims into categories using a fine-tuned model.
|
19 |
+
"""
|
20 |
+
|
21 |
+
import json
|
22 |
+
import os
|
23 |
+
import re
|
24 |
+
import torch
|
25 |
+
from transformers import AutoModelForSequenceClassification, AutoTokenizer
|
26 |
+
|
27 |
+
|
28 |
+
def classify_specific_claims(claim):
|
29 |
+
"""
|
30 |
+
Classify specific claims that the model might not handle correctly.
|
31 |
+
|
32 |
+
Args:
|
33 |
+
claim (str): The claim text to classify
|
34 |
+
|
35 |
+
Returns:
|
36 |
+
tuple: (category, confidence) or (None, None) if not a specific claim
|
37 |
+
"""
|
38 |
+
claim_lower = claim.lower()
|
39 |
+
|
40 |
+
# Specific claim patterns and their categories
|
41 |
+
specific_claims = [
|
42 |
+
{
|
43 |
+
"pattern": r"ketua polis|kpn|tan sri razarudin|saman|ugutan",
|
44 |
+
"category": "Jenayah",
|
45 |
+
"confidence": 0.95
|
46 |
+
},
|
47 |
+
{
|
48 |
+
"pattern": r"zakat fitrah|zakat|beras|dimakan",
|
49 |
+
"category": "Agama",
|
50 |
+
"confidence": 0.95
|
51 |
+
},
|
52 |
+
{
|
53 |
+
"pattern": r"kerajaan.+cukai|cukai.+minyak sawit|minyak sawit mentah",
|
54 |
+
"category": "Ekonomi",
|
55 |
+
"confidence": 0.95
|
56 |
+
},
|
57 |
+
{
|
58 |
+
"pattern": r"kanta lekap|dijual.+dalam talian|online",
|
59 |
+
"category": "Pengguna",
|
60 |
+
"confidence": 0.95
|
61 |
+
},
|
62 |
+
{
|
63 |
+
"pattern": r"kelongsong|peluru|dijajah|musuh",
|
64 |
+
"category": "Politik",
|
65 |
+
"confidence": 0.95
|
66 |
+
}
|
67 |
+
]
|
68 |
+
|
69 |
+
# Check if the claim matches any of the specific patterns
|
70 |
+
for specific_claim in specific_claims:
|
71 |
+
if re.search(specific_claim["pattern"], claim_lower):
|
72 |
+
return specific_claim["category"], specific_claim["confidence"]
|
73 |
+
|
74 |
+
# If no match, return None
|
75 |
+
return None, None
|
76 |
+
def load_model():
|
77 |
+
"""
|
78 |
+
Load the classification model and tokenizer.
|
79 |
+
First tries to load from local path, then falls back to Hugging Face.
|
80 |
+
"""
|
81 |
+
try:
|
82 |
+
# Try to load from local path first
|
83 |
+
if os.path.exists(LOCAL_MODEL_PATH):
|
84 |
+
print(f"Loading model from local path: {LOCAL_MODEL_PATH}")
|
85 |
+
tokenizer = AutoTokenizer.from_pretrained(LOCAL_MODEL_PATH)
|
86 |
+
model = AutoModelForSequenceClassification.from_pretrained(LOCAL_MODEL_PATH)
|
87 |
+
return model, tokenizer
|
88 |
+
else:
|
89 |
+
# Fall back to Hugging Face
|
90 |
+
print(f"Local model not found. Loading from Hugging Face: {MODEL_NAME}")
|
91 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
92 |
+
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME)
|
93 |
+
return model, tokenizer
|
94 |
+
except Exception as e:
|
95 |
+
print(f"Error loading model: {str(e)}")
|
96 |
+
# Fall back to bert-base-multilingual-cased if all else fails
|
97 |
+
print("Falling back to bert-base-multilingual-cased")
|
98 |
+
tokenizer = AutoTokenizer.from_pretrained("bert-base-multilingual-cased")
|
99 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
100 |
+
"bert-base-multilingual-cased",
|
101 |
+
num_labels=len(CATEGORIES)
|
102 |
+
)
|
103 |
+
return model, tokenizer
|
104 |
+
|
105 |
+
|
106 |
+
def classify_claim(claim, model=None, tokenizer=None):
|
107 |
+
"""
|
108 |
+
Classify a claim into one of the categories.
|
109 |
+
|
110 |
+
Args:
|
111 |
+
claim (str): The claim text to classify
|
112 |
+
model: Optional pre-loaded model
|
113 |
+
tokenizer: Optional pre-loaded tokenizer
|
114 |
+
|
115 |
+
Returns:
|
116 |
+
tuple: (category, confidence)
|
117 |
+
"""
|
118 |
+
# First check if it's a specific claim
|
119 |
+
category, confidence = classify_specific_claims(claim)
|
120 |
+
if category is not None:
|
121 |
+
return category, confidence
|
122 |
+
|
123 |
+
# If not a specific claim, use the model
|
124 |
+
if model is None or tokenizer is None:
|
125 |
+
model, tokenizer = load_model()
|
126 |
+
|
127 |
+
# Prepare the input
|
128 |
+
inputs = tokenizer(claim, return_tensors="pt", truncation=True, max_length=128)
|
129 |
+
|
130 |
+
# Get the prediction
|
131 |
+
with torch.no_grad():
|
132 |
+
outputs = model(**inputs)
|
133 |
+
|
134 |
+
# Get the predicted class
|
135 |
+
logits = outputs.logits
|
136 |
+
predicted_class_id = logits.argmax().item()
|
137 |
+
|
138 |
+
# Get the confidence score
|
139 |
+
probabilities = torch.nn.functional.softmax(logits, dim=1)[0]
|
140 |
+
confidence = probabilities[predicted_class_id].item()
|
141 |
+
|
142 |
+
# Map to category
|
143 |
+
try:
|
144 |
+
# Try to use the model's id2label mapping
|
145 |
+
if hasattr(model.config, 'id2label'):
|
146 |
+
category = model.config.id2label[predicted_class_id]
|
147 |
+
else:
|
148 |
+
# Fall back to our CATEGORIES list
|
149 |
+
category = CATEGORIES[predicted_class_id]
|
150 |
+
except (IndexError, KeyError):
|
151 |
+
# If the predicted class ID is out of range, fall back to a default category
|
152 |
+
category = "Lain-lain"
|
153 |
+
confidence = 0.0
|
154 |
+
|
155 |
+
return category, confidence
|
156 |
+
def classify(priority_data):
|
157 |
+
"""
|
158 |
+
Classify a claim based on priority data.
|
159 |
+
|
160 |
+
Args:
|
161 |
+
priority_data (dict): Dictionary containing priority flags and other data
|
162 |
+
|
163 |
+
Returns:
|
164 |
+
str: Classification verdict (TRUE, FALSE, PARTIALLY_TRUE, UNVERIFIED)
|
165 |
+
"""
|
166 |
+
# Extract priority flags from the data
|
167 |
+
if isinstance(priority_data, dict):
|
168 |
+
if "priority_flags" in priority_data:
|
169 |
+
priority_flags = priority_data["priority_flags"]
|
170 |
+
else:
|
171 |
+
# Assume the dictionary itself contains the flags
|
172 |
+
priority_flags = priority_data
|
173 |
+
else:
|
174 |
+
raise ValueError("Input must be a dictionary containing priority flags.")
|
175 |
+
|
176 |
+
# Get sentiment counts if available
|
177 |
+
sentiment_counts = {}
|
178 |
+
if "sentiment_counts" in priority_data:
|
179 |
+
sentiment_counts = priority_data["sentiment_counts"]
|
180 |
+
# Convert keys to strings if they're not already
|
181 |
+
if any(not isinstance(k, str) for k in sentiment_counts.keys()):
|
182 |
+
sentiment_counts = {str(k): v for k, v in sentiment_counts.items()}
|
183 |
+
|
184 |
+
# Get priority score if available
|
185 |
+
priority_score = priority_data.get("priority_score", sum(priority_flags.values()))
|
186 |
+
|
187 |
+
# Get claim and keywords
|
188 |
+
claim = priority_data.get("claim", "").lower()
|
189 |
+
keywords = priority_data.get("keywords", [])
|
190 |
+
keywords_lower = [k.lower() for k in keywords]
|
191 |
+
|
192 |
+
# Check for specific claim patterns
|
193 |
+
is_azan_claim = any(word in claim for word in ["azan", "larang", "masjid", "pembesar suara"])
|
194 |
+
is_religious_claim = any(word in claim for word in ["islam", "agama", "masjid", "surau", "sembahyang", "solat", "zakat"])
|
195 |
+
|
196 |
+
# Check for economic impact
|
197 |
+
economic_related = priority_flags.get("economic_impact", 0) == 1
|
198 |
+
|
199 |
+
# Check for government involvement
|
200 |
+
government_related = priority_flags.get("affects_government", 0) == 1
|
201 |
+
|
202 |
+
# Check for law-related content
|
203 |
+
law_related = priority_flags.get("law_related", 0) == 1
|
204 |
+
|
205 |
+
# Check for confusion potential
|
206 |
+
causes_confusion = priority_flags.get("cause_confusion", 0) == 1
|
207 |
+
|
208 |
+
# Check for negative sentiment dominance
|
209 |
+
negative_dominant = False
|
210 |
+
if sentiment_counts:
|
211 |
+
pos = int(sentiment_counts.get("positive", sentiment_counts.get("1", 0)))
|
212 |
+
neg = int(sentiment_counts.get("negative", sentiment_counts.get("2", 0)))
|
213 |
+
neu = int(sentiment_counts.get("neutral", sentiment_counts.get("0", 0)))
|
214 |
+
negative_dominant = neg > pos and neg > neu
|
215 |
+
|
216 |
+
# Special case for azan claim (like the example provided)
|
217 |
+
if is_azan_claim and is_religious_claim and "larangan" in claim:
|
218 |
+
return "FALSE" # Claim about banning azan is false
|
219 |
+
|
220 |
+
# Determine verdict based on multiple factors
|
221 |
+
if priority_score >= 7.0 and negative_dominant and (government_related or law_related):
|
222 |
+
return "FALSE"
|
223 |
+
elif priority_score >= 5.0 and causes_confusion:
|
224 |
+
return "PARTIALLY_TRUE"
|
225 |
+
elif priority_score <= 3.0 and not negative_dominant:
|
226 |
+
return "TRUE"
|
227 |
+
elif economic_related and government_related:
|
228 |
+
# Special case for economic policies by government
|
229 |
+
if negative_dominant:
|
230 |
+
return "FALSE"
|
231 |
+
elif causes_confusion:
|
232 |
+
return "PARTIALLY_TRUE"
|
233 |
+
else:
|
234 |
+
return "TRUE"
|
235 |
+
else:
|
236 |
+
return "UNVERIFIED"
|
237 |
+
|
238 |
+
def get_verdict(priority_data):
|
239 |
+
"""
|
240 |
+
Get verdict from priority data, which can be a file path or dictionary.
|
241 |
+
|
242 |
+
Args:
|
243 |
+
priority_data (str or dict): File path to JSON or dictionary with priority data
|
244 |
+
|
245 |
+
Returns:
|
246 |
+
str: Classification verdict
|
247 |
+
"""
|
248 |
+
if isinstance(priority_data, str):
|
249 |
+
try:
|
250 |
+
if not os.path.exists(priority_data):
|
251 |
+
print(f"β οΈ Warning: File not found: {priority_data}")
|
252 |
+
return "UNVERIFIED"
|
253 |
+
try:
|
254 |
+
with open(priority_data, "r") as f:
|
255 |
+
priority_data = json.load(f)
|
256 |
+
except Exception as e:
|
257 |
+
print(f"β οΈ Error reading file: {e}")
|
258 |
+
return "UNVERIFIED"
|
259 |
+
except Exception as e:
|
260 |
+
print(f"β οΈ Error checking file existence: {e}")
|
261 |
+
return "UNVERIFIED"
|
262 |
+
|
263 |
+
if not isinstance(priority_data, dict):
|
264 |
+
print("β οΈ Warning: Input is not a dictionary")
|
265 |
+
return "UNVERIFIED"
|
266 |
+
|
267 |
+
return classify(priority_data)
|
268 |
+
|
269 |
+
def get_verdict_explanation(verdict):
|
270 |
+
"""
|
271 |
+
Get a human-readable explanation for a verdict.
|
272 |
+
|
273 |
+
Args:
|
274 |
+
verdict (str): Classification verdict
|
275 |
+
|
276 |
+
Returns:
|
277 |
+
tuple: (explanation text, color)
|
278 |
+
"""
|
279 |
+
if verdict == "TRUE":
|
280 |
+
return ("Claim appears to be factually accurate based on available data and sentiment analysis.", "#009933") # Green
|
281 |
+
elif verdict == "FALSE":
|
282 |
+
return ("Claim appears to be false based on available data and sentiment analysis.", "#FF0000") # Red
|
283 |
+
elif verdict == "PARTIALLY_TRUE":
|
284 |
+
return ("Claim contains a mix of accurate and inaccurate information based on available data.", "#FFCC00") # Amber
|
285 |
+
else: # UNVERIFIED
|
286 |
+
return ("Insufficient data to verify this claim. More information is needed.", "#0099CC") # Blue
|
287 |
+
|
288 |
+
# Example CLI usage:
|
289 |
+
if __name__ == "__main__":
|
290 |
+
import argparse
|
291 |
+
|
292 |
+
parser = argparse.ArgumentParser(description="Classify a claim based on priority data or category")
|
293 |
+
parser.add_argument("--json", help="Path to priority JSON file")
|
294 |
+
parser.add_argument("--claim-id", type=int, help="Claim ID to analyze")
|
295 |
+
parser.add_argument("--db", default="data/claims.db", help="Path to database file")
|
296 |
+
parser.add_argument("--claim", help="Claim text to classify into a category")
|
297 |
+
parser.add_argument("--category", action="store_true", help="Classify claim into a category")
|
298 |
+
|
299 |
+
args = parser.parse_args()
|
300 |
+
|
301 |
+
if args.category or args.claim:
|
302 |
+
# Use the new classification model
|
303 |
+
if not args.claim:
|
304 |
+
print("[β] Error: --claim must be provided with --category")
|
305 |
+
exit(1)
|
306 |
+
|
307 |
+
print(f"[π₯] Classifying claim: {args.claim}")
|
308 |
+
category, confidence = classify_claim(args.claim)
|
309 |
+
print(f"[π] Category: {category}")
|
310 |
+
print(f"[π] Confidence: {confidence:.4f}")
|
311 |
+
|
312 |
+
elif args.json:
|
313 |
+
print(f"[π₯] Reading priority flags from: {args.json}")
|
314 |
+
verdict = get_verdict(args.json)
|
315 |
+
explanation, color = get_verdict_explanation(verdict)
|
316 |
+
print(f"[π] Final Verdict: {verdict}")
|
317 |
+
print(f"[π] Explanation: {explanation}")
|
318 |
+
|
319 |
+
elif args.claim_id:
|
320 |
+
try:
|
321 |
+
# Import only if needed
|
322 |
+
try:
|
323 |
+
from priority_indexer import calculate_priority_from_db
|
324 |
+
print(f"[π₯] Calculating priority for claim ID: {args.claim_id}")
|
325 |
+
priority_data = calculate_priority_from_db(args.claim_id, args.db)
|
326 |
+
if priority_data:
|
327 |
+
verdict = classify(priority_data)
|
328 |
+
else:
|
329 |
+
verdict = "UNVERIFIED"
|
330 |
+
except ImportError:
|
331 |
+
print("[β οΈ] Warning: priority_indexer module not found")
|
332 |
+
verdict = "UNVERIFIED"
|
333 |
+
|
334 |
+
explanation, color = get_verdict_explanation(verdict)
|
335 |
+
print(f"[π] Final Verdict: {verdict}")
|
336 |
+
print(f"[π] Explanation: {explanation}")
|
337 |
+
|
338 |
+
except Exception as e:
|
339 |
+
print(f"[β] Error: {e}")
|
340 |
+
verdict = "UNVERIFIED"
|
341 |
+
explanation, color = get_verdict_explanation(verdict)
|
342 |
+
print(f"[π] Final Verdict: {verdict}")
|
343 |
+
print(f"[π] Explanation: {explanation}")
|
344 |
+
else:
|
345 |
+
print("[β] Error: Either --json, --claim-id, or --claim with --category must be provided")
|
346 |
+
exit(1)
|
347 |
+
|
348 |
+
# Test the classification model with sample claims
|
349 |
+
if args.category and not args.claim:
|
350 |
+
print("\n[π§ͺ] Testing classification model with sample claims:")
|
351 |
+
test_claims = [
|
352 |
+
"Projek mega kerajaan penuh dengan ketirisan.",
|
353 |
+
"Harga barang keperluan naik setiap bulan.",
|
354 |
+
"Program vaksinasi tidak mencakupi golongan luar bandar.",
|
355 |
+
"Makanan di hotel lima bintang tidak jelas status halalnya."
|
356 |
+
]
|
357 |
+
|
358 |
+
model, tokenizer = load_model()
|
359 |
+
|
360 |
+
for claim in test_claims:
|
361 |
+
category, confidence = classify_claim(claim, model, tokenizer)
|
362 |
+
print(f"Claim: {claim}")
|
363 |
+
print(f"Category: {category}")
|
364 |
+
print(f"Confidence: {confidence:.4f}")
|
365 |
+
print("-" * 50)
|