Spaces:
Sleeping
Sleeping
File size: 18,790 Bytes
070061f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 |
# NLU module for sentence-transformers-based semantic similarity and intent extraction
import pandas as pd
import os
import numpy as np
from constraints import L_SUPPORT_QUESTIONS_IDS, INTENT_TO_XAI_METHOD
try:
from sentence_transformers import SentenceTransformer
SENTENCE_TRANSFORMERS_AVAILABLE = True
except ImportError:
SentenceTransformer = None
SENTENCE_TRANSFORMERS_AVAILABLE = False
try:
from simcse import SimCSE
SIMCSE_AVAILABLE = True
except ImportError:
SimCSE = None
SIMCSE_AVAILABLE = False
class NLU:
def __init__(self, model_type="sentence_transformers", model_path=None):
self.model_type = model_type
self.df = pd.read_csv(os.path.join(os.path.dirname(__file__), '..', 'data_questions', 'Median_4.csv'), index_col=0).drop_duplicates()
self.questions = list(self.df['Question'])
# Prefer sentence-transformers; use GPU if available, otherwise CPU (Streamlit Cloud has no GPU)
if model_type == "sentence_transformers":
if not SENTENCE_TRANSFORMERS_AVAILABLE:
print("β οΈ sentence-transformers not available, trying SimCSE...")
self.model_type = "simcse"
else:
try:
import torch
device = "cuda" if torch.cuda.is_available() else "cpu"
except Exception:
device = "cpu"
# Lightweight, fast model for semantic similarity
self.model = SentenceTransformer('all-MiniLM-L6-v2', device=device)
print(f"β
Loaded sentence-transformers model on {device}")
# Pre-compute embeddings for all questions
self.question_embeddings = self.model.encode(self.questions, convert_to_numpy=True, show_progress_bar=False)
print(f"β
Pre-computed embeddings for {len(self.questions)} questions")
# Optional SimCSE fallback for legacy envs
if self.model_type == "simcse" or (model_type == "sentence_transformers" and not SENTENCE_TRANSFORMERS_AVAILABLE):
if not SIMCSE_AVAILABLE:
print("β οΈ SimCSE not available, falling back to simple keyword matching")
self.model_type = "fallback"
self.model = None
else:
self.model = SimCSE("princeton-nlp/sup-simcse-roberta-large")
self.model.build_index(self.questions)
self.model_type = "simcse"
elif model_type == "fallback":
self.model = None
elif self.model_type not in {"sentence_transformers", "simcse", "fallback"}:
raise ValueError(f"Unsupported NLU model type: {model_type}. Supported: 'sentence_transformers', 'simcse', 'fallback'")
def classify_intent(self, user_input, top_k=5):
# Dynamic, model-driven intent extraction
# Fast keyword heuristics ensure clear phrases immediately map to an XAI method
try:
text = (user_input or "").lower()
# Heuristics for common phrasing
rule_keywords = ["rule-based", "rule based", "rules", "conditions", "if then", "anchor"]
shap_keywords = ["feature", "importance", "impact", "influence", "contribute", "shap", "why", "explain", "decision", "factors", "affected"]
dice_keywords = ["what if", "counterfactual", "change", "modify", "different", "should i", "how to get"]
if any(k in text for k in rule_keywords):
return {
'intent': 'anchor',
'label': None,
'confidence': 0.95,
'matched_question': "Provide a simple rule-based explanation for this decision."
}, 0.95, []
# Explicitly detect single-word 'why' queries too
if text.strip() == "why" or any(k in text for k in shap_keywords):
return {
'intent': 'shap',
'label': None,
'confidence': 0.9,
'matched_question': "Which features were most important for this prediction?"
}, 0.9, []
if any(k in text for k in dice_keywords):
return {
'intent': 'dice',
'label': None,
'confidence': 0.9,
'matched_question': "How should the instance be changed to get a different prediction?"
}, 0.9, []
except Exception:
pass
# sentence-transformers path
if self.model_type == "sentence_transformers" and hasattr(self, 'question_embeddings'):
try:
query_emb = self.model.encode([user_input], convert_to_numpy=True, show_progress_bar=False)[0]
# Cosine similarity
q_norm = np.linalg.norm(self.question_embeddings, axis=1) + 1e-12
u_norm = np.linalg.norm(query_emb) + 1e-12
sims = (self.question_embeddings @ query_emb) / (q_norm * u_norm)
# Top-k indices
top_idx = np.argsort(-sims)[:top_k]
match_question = self.questions[top_idx[0]]
score = float(sims[top_idx[0]])
label = self.df.iloc[top_idx[0]]['Label']
xai_method = self.map_label_to_xai_method(label)
suggestions = [self.questions[i] for i in top_idx]
return {
'intent': xai_method,
'label': label,
'confidence': score,
'matched_question': match_question
}, score, suggestions
except Exception as e:
print(f"sentence-transformers classify failed: {e}")
# Legacy SimCSE path
if self.model_type == "simcse" and self.model is not None:
# Always get top matches without initial threshold filtering
match_results = self.model.search(user_input, threshold=0, top_k=top_k)
if len(match_results) > 0:
match_question, score = match_results[0]
# Get the label for the matched question
label = self.df.query('Question == @match_question')['Label'].iloc[0]
# Map label to XAI method if supported
xai_method = self.map_label_to_xai_method(label)
# Normalize confidence score to 0-1 range for consistency
# SimCSE scores can be very high, so we'll use relative confidence
normalized_confidence = min(1.0, score / 1e20) if score > 1 else score
# Always return the best match but indicate confidence level
return {
'intent': xai_method,
'label': label,
'confidence': normalized_confidence,
'matched_question': match_question
}, normalized_confidence, []
# Fallback to simple keyword matching when SimCSE is not available
elif self.model_type == "fallback" or self.model is None:
return self._fallback_classify_intent(user_input, top_k)
# No matches found at all
return 'unknown', 0.0, []
else:
return 'unknown', 0.0, []
def match(self, user_input, features=None, prediction=None, current_instance=None, labels=None):
"""Hybrid approach: Fuzzy first (primary), Intent classifier fallback"""
# PRIMARY: Try fuzzy matching first (fast and reliable)
fuzzy_result = self._fuzzy_match_fallback(user_input)
if fuzzy_result != "unknown":
print(f"π€ Fuzzy match (primary): {fuzzy_result}")
return fuzzy_result
# FALLBACK 1: Try intent classifier (65% accuracy)
intent_result = self._classify_with_intent_classifier(user_input)
if intent_result != "unknown":
print(f"π§ Intent classifier (fallback): {intent_result}")
return intent_result
# FALLBACK 2: Try embedding search if available (ST first, then SimCSE)
if self.model_type == "sentence_transformers" and hasattr(self, 'question_embeddings'):
try:
query_emb = self.model.encode([user_input], convert_to_numpy=True, show_progress_bar=False)[0]
q_norm = np.linalg.norm(self.question_embeddings, axis=1) + 1e-12
u_norm = np.linalg.norm(query_emb) + 1e-12
sims = (self.question_embeddings @ query_emb) / (q_norm * u_norm)
best_idx = int(np.argmax(sims))
match_question = self.questions[best_idx]
print(f"π ST match (last resort): {match_question}")
return match_question
except Exception as e:
print(f"ST search failed: {e}")
if hasattr(self, 'model') and self.model_type == "simcse" and self.model is not None:
try:
threshold = 0.6
match_results = self.model.search(user_input, threshold=threshold)
if len(match_results) > 0:
match_question, score = match_results[0]
print(f"π SimCSE match (last resort): {match_question}")
return match_question
else:
# Try with no threshold
match_results = self.model.search(user_input, threshold=0, top_k=5)
if len(match_results) > 0:
match_question, score = match_results[0]
print(f"π SimCSE fallback: {match_question}")
return match_question
except Exception as e:
print(f"SimCSE search failed: {e}")
print(f"β No match found for: '{user_input}'")
return "unknown"
def _fuzzy_match_fallback(self, user_input):
"""Fallback fuzzy matching using simple string similarity"""
try:
from difflib import SequenceMatcher
user_lower = user_input.lower()
best_match = None
best_score = 0
# Define key patterns for different XAI methods
shap_patterns = [
"feature", "important", "impact", "contribute", "influence", "matter", "weigh", "explain", "why"
]
dice_patterns = [
"change", "different", "modify", "counterfact", "should", "what if", "approved", "denied"
]
anchor_patterns = [
"rule", "condition", "guarantee", "necessary", "sufficient", "always", "simple"
]
# Check for pattern matches
if any(pattern in user_lower for pattern in shap_patterns):
# Return a representative SHAP question
return "What features of this instance lead to the system's prediction?"
elif any(pattern in user_lower for pattern in dice_patterns):
# Return a representative DiCE question
return "How should the instance be changed to get a different (better or worse) prediction?"
elif any(pattern in user_lower for pattern in anchor_patterns):
# Return a representative Anchor question
return "What is the minimum requirement for the prediction to stay the same?"
# If no patterns match, try fuzzy string matching with dataset questions
for _, row in self.df.iterrows():
question = row['Question']
similarity = SequenceMatcher(None, user_lower, question.lower()).ratio()
if similarity > best_score:
best_score = similarity
best_match = question
# Return best match if similarity is reasonable
if best_score > 0.4: # 40% similarity threshold
return best_match
except Exception as e:
print(f"Fuzzy matching failed: {e}")
return "unknown"
def get_question_suggestions(self, match_results):
"""Extract question suggestions from match results"""
suggestions = []
for question, _ in match_results:
if len(suggestions) < 5: # Limit to 5 suggestions
suggestions.append(question)
return suggestions
def map_label_to_xai_method(self, label):
"""Map question label to appropriate XAI method (adopted from XAgent logic)"""
from constraints import L_SHAP_QUESTION_IDS, L_DICE_QUESTION_IDS, L_ANCHOR_QUESTION_IDS
if label in L_SHAP_QUESTION_IDS:
return "shap"
elif label in L_DICE_QUESTION_IDS:
return "dice"
elif label in L_ANCHOR_QUESTION_IDS:
return "anchor"
else:
return "general"
def replace_information(self, question, features=None, prediction=None, current_instance=None, labels=None):
"""Replace template variables in questions (adopted from XAgent)"""
if features and "{X}" in question:
feature_str = f"{{{features[0]},{features[1]}, ...}}" if len(features) > 1 else f"{features[0]}"
question = question.replace("{X}", feature_str)
if prediction and "{P}" in question:
question = question.replace("{P}", str(prediction))
if labels and prediction and "{Q}" in question:
other_labels = [label for label in labels if str(label) != str(prediction)]
question = question.replace("{Q}", str(other_labels))
return question
def _classify_with_intent_classifier(self, user_input):
"""Use the trained intent classifier (65% accuracy) as fallback"""
try:
# Try to load intent classifier if not already loaded
if not hasattr(self, 'intent_classifier') or self.intent_classifier is None:
self._load_intent_classifier()
if self.intent_classifier is None:
return "unknown"
# Generate embedding for user input
embedding = self.intent_simcse.encode([user_input])
# Convert to tensor
import torch
import numpy as np
embedding_tensor = torch.FloatTensor(embedding)
# Get classifier prediction
with torch.no_grad():
outputs = self.intent_classifier(embedding_tensor)
probabilities = outputs[0].numpy()
# Get the class with highest probability
predicted_class_idx = np.argmax(probabilities)
confidence = probabilities[predicted_class_idx]
# Use lower threshold since this is fallback
if confidence >= 0.3: # Lower threshold for fallback
# Convert back to intent
predicted_intent = self.intent_label_encoder.inverse_transform([predicted_class_idx])[0]
# Map intent to representative question
if predicted_intent == 'shap':
return "What features of this instance lead to the system's prediction?"
elif predicted_intent == 'dice':
return "How should the instance be changed to get a different (better or worse) prediction?"
elif predicted_intent == 'anchor':
return "What is the minimum requirement for the prediction to stay the same?"
# Don't return anything for 'other' - let it fall through
except Exception as e:
print(f"Intent classifier failed: {e}")
return "unknown"
def _load_intent_classifier(self):
"""Load the trained intent classifier (65% accuracy model)"""
try:
import torch
import torch.nn as nn
import pickle
import numpy as np
from simcse import SimCSE
# Define the classifier architecture (matching the training script)
class IntentClassifier(nn.Module):
def __init__(self, input_dim, hidden_dim, num_classes=4):
super(IntentClassifier, self).__init__()
self.network = nn.Sequential(
nn.Linear(input_dim, hidden_dim),
nn.ReLU(),
nn.Dropout(0.2),
nn.Linear(hidden_dim, hidden_dim // 2),
nn.ReLU(),
nn.Dropout(0.2),
nn.Linear(hidden_dim // 2, num_classes),
nn.Softmax(dim=1)
)
def forward(self, x):
return self.network(x)
# Load metadata
with open('models/intent_classifier_metadata.pkl', 'rb') as f:
metadata = pickle.load(f)
# Load label encoder
with open('models/intent_label_encoder.pkl', 'rb') as f:
self.intent_label_encoder = pickle.load(f)
# Initialize and load classifier
self.intent_classifier = IntentClassifier(
metadata['input_dim'],
metadata['hidden_dim'],
metadata['num_classes']
)
self.intent_classifier.load_state_dict(torch.load('models/intent_classifier_best.pth', map_location='cpu'))
self.intent_classifier.eval()
# Initialize SimCSE for embedding generation
self.intent_simcse = SimCSE("princeton-nlp/sup-simcse-roberta-large")
print(f"β
Loaded intent classifier (accuracy: {metadata.get('best_accuracy', 'unknown'):.4f})")
except Exception as e:
print(f"β οΈ Could not load intent classifier: {e}")
self.intent_classifier = None
self.intent_label_encoder = None
self.intent_simcse = None
|