File size: 8,621 Bytes
31fda96 0117df3 31fda96 0117df3 31fda96 0117df3 88da32f 31fda96 88da32f 31fda96 88da32f 31fda96 0117df3 31fda96 0117df3 31fda96 0117df3 31fda96 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 | from __future__ import annotations
from dataclasses import dataclass
from functools import lru_cache
import logging
import random
from typing import Any
import nltk
import numpy as np
from scipy.sparse import csr_matrix, hstack
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from features.text_classifier.model_loader import load_model
logger = logging.getLogger(__name__)
for resource in ("tokenizers/punkt", "tokenizers/punkt_tab"):
try:
nltk.data.find(resource)
except LookupError:
nltk.download(resource.split("/")[-1], quiet=True)
try:
import textstat
except ImportError:
textstat = None
@dataclass
class SentenceBlendConfig:
sentence_blend_weight: float = 0.70
sentence_to_doc_bias: float = 0.35
max_sentence_blend_weight: float = 0.90
max_sentence_to_doc_bias: float = 0.80
random_deviation_pct: float = 2.0
class PerplexityCalculator:
"""Lazy-loaded perplexity calculator for distilgpt2."""
def __init__(self, model_name: str = "distilgpt2"):
self.model_name = model_name
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self._tokenizer = None
self._model = None
def _load(self) -> None:
if self._model is not None and self._tokenizer is not None:
return
logger.info("Loading perplexity model: %s", self.model_name)
self._tokenizer = AutoTokenizer.from_pretrained(self.model_name)
self._model = AutoModelForCausalLM.from_pretrained(self.model_name).to(self.device)
self._model.eval()
logger.info("Perplexity model loaded on %s", self.device)
def calculate(self, text: str, max_length: int = 512) -> float:
try:
self._load()
encodings = self._tokenizer(
text,
return_tensors="pt",
truncation=True,
max_length=max_length,
)
input_ids = encodings.input_ids.to(self.device)
with torch.no_grad():
outputs = self._model(input_ids, labels=input_ids)
loss = outputs.loss
perplexity = torch.exp(loss).item()
return min(float(perplexity), 10000.0)
except Exception as exc:
logger.warning("Perplexity fallback used due to error: %s", exc)
return 100.0
_perplexity_calc = PerplexityCalculator()
@lru_cache(maxsize=20000)
def _cached_perplexity(cleaned_text: str) -> float:
return _perplexity_calc.calculate(cleaned_text)
@lru_cache(maxsize=1)
def _get_model_artifacts() -> tuple[Any, Any, Any, Any, list[str], dict[str, Any]]:
return load_model()
def normalize_text(text: str) -> str:
return " ".join(str(text).split()).strip()
def split_into_sentences(text: str) -> list[str]:
cleaned = normalize_text(text)
if not cleaned:
return []
sentences = [s.strip() for s in nltk.sent_tokenize(cleaned) if s.strip()]
return sentences if sentences else [cleaned]
def extract_burstiness_features(text: str) -> dict[str, float]:
sentences = split_into_sentences(text)
if not sentences:
return {
"burst_mean": 0.0,
"burst_std": 0.0,
"burst_max": 0.0,
"burst_min": 0.0,
"burst_range": 0.0,
}
lengths = np.array([len(s.split()) for s in sentences], dtype=float)
return {
"burst_mean": float(np.mean(lengths)),
"burst_std": float(np.std(lengths)),
"burst_max": float(np.max(lengths)),
"burst_min": float(np.min(lengths)),
"burst_range": float(np.max(lengths) - np.min(lengths)),
}
def extract_stylometry_features(text: str) -> dict[str, float]:
words = text.split()
num_words = len(words)
num_chars = len(text)
num_sentences = max(len(split_into_sentences(text)), 1)
avg_word_len = float(np.mean([len(w) for w in words])) if words else 0.0
avg_sent_len = float(num_words / num_sentences)
unique_words = len(set(words))
lexical_diversity = float(unique_words / num_words) if num_words > 0 else 0.0
num_punct = sum(1 for c in text if c in ".,!?;:")
punct_ratio = float(num_punct / num_chars) if num_chars > 0 else 0.0
num_caps = sum(1 for c in text if c.isupper())
caps_ratio = float(num_caps / num_chars) if num_chars > 0 else 0.0
if textstat is not None:
try:
flesch_reading = float(textstat.flesch_reading_ease(text))
flesch_grade = float(textstat.flesch_kincaid_grade(text))
except Exception:
flesch_reading = 50.0
flesch_grade = 8.0
else:
flesch_reading = 50.0
flesch_grade = 8.0
return {
"num_words": float(num_words),
"num_chars": float(num_chars),
"num_sentences": float(num_sentences),
"avg_word_len": avg_word_len,
"avg_sent_len": avg_sent_len,
"lexical_diversity": lexical_diversity,
"punct_ratio": punct_ratio,
"caps_ratio": caps_ratio,
"flesch_reading": flesch_reading,
"flesch_grade": flesch_grade,
}
def extract_all_features(text: str, calc_perplexity: bool = True) -> dict[str, float]:
cleaned = normalize_text(text)
features: dict[str, float] = {}
if calc_perplexity:
features["perplexity"] = _cached_perplexity(cleaned)
else:
features["perplexity"] = 100.0
features.update(extract_burstiness_features(cleaned))
features.update(extract_stylometry_features(cleaned))
return features
def _predict_ai_probability(text: str) -> tuple[float, float]:
(
loaded_classifier,
loaded_scaler,
loaded_word_vectorizer,
loaded_char_vectorizer,
loaded_features,
loaded_metadata,
) = _get_model_artifacts()
calc_perplexity = bool(loaded_metadata.get("num_engineered_features", 0) > 0)
features = extract_all_features(text, calc_perplexity=calc_perplexity)
feature_vector = np.array([features[name] for name in loaded_features], dtype=float).reshape(1, -1)
feature_scaled = loaded_scaler.transform(feature_vector)
word_vec = loaded_word_vectorizer.transform([text])
char_vec = loaded_char_vectorizer.transform([text])
num_vec = csr_matrix(feature_scaled)
hybrid_vec = hstack([word_vec, char_vec, num_vec], format="csr")
if hasattr(loaded_classifier, "predict_proba"):
proba = loaded_classifier.predict_proba(hybrid_vec)[0]
ai_prob = float(proba[1])
else:
score = float(loaded_classifier.decision_function(hybrid_vec)[0])
ai_prob = float(1.0 / (1.0 + np.exp(-score)))
perplexity = float(features.get("perplexity", 100.0))
return ai_prob, perplexity
def classify_text(text: str) -> tuple[str, float, float]:
"""Return (label, perplexity, ai_likelihood_percent)."""
cleaned = normalize_text(text)
if not cleaned:
raise ValueError("Input text is empty")
ai_prob, perplexity = _predict_ai_probability(cleaned)
ai_likelihood = round(ai_prob * 100.0, 2)
label = "AI" if ai_likelihood >= 50.0 else "Human"
return label, perplexity, ai_likelihood
def analyze_text_with_sentences(
text: str,
) -> dict[str, Any]:
text = normalize_text(text)
overall_classification, overall_perplexity, overall_ai_likelihood = classify_text(text)
sentences = split_into_sentences(text)
if not sentences:
raise ValueError("Input text contains no valid sentences")
# do the per-sentence analysis
sentence_results = []
for sentence in sentences:
try:
label, perplexity, ai_likelihood = classify_text(sentence)
sentence_results.append(
{
"sentence": sentence,
"label": label,
"perplexity": perplexity,
"ai_likelihood": ai_likelihood,
}
)
except Exception as exc:
logger.warning("Error analyzing sentence: %s", exc)
sentence_results.append(
{
"sentence": sentence,
"label": "Error",
"perplexity": None,
"ai_likelihood": None,
}
)
return{
"sentences": sentence_results,
"summary": {
"overall": {
"label": overall_classification,
"perplexity": overall_perplexity,
"ai_likelihood": overall_ai_likelihood,
}
},
}
|