Spaces:
Running on Zero
Running on Zero
File size: 4,750 Bytes
90fa9aa | 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 | import sys
import numpy as np
from pathlib import Path
sys.path.append(str(Path(__file__).resolve().parent.parent.parent))
from src.models.text_classifier import LinguisticStressClassifier
class TextExplainerLIME:
"""
Token-Level Explainability Engine for Text Stress Classification.
Inspired by LIME (Local Interpretable Model-agnostic Explanations) as demonstrated in Paper #14.
Computes exact contribution scores for individual words by analyzing perturbation impact
and TF-IDF vocabulary weights.
"""
def __init__(self, text_classifier=None):
self.text_classifier = text_classifier or LinguisticStressClassifier()
if not self.text_classifier.is_fitted:
try:
self.text_classifier.load_model()
except Exception:
pass
def explain_instance(self, text, top_k=8):
"""
Calculates word importance weights for the given text.
Returns both raw scores and an HTML color-highlighted representation for UI rendering.
"""
words = text.split()
if not words:
return {"word_scores": [], "html_highlighted": "", "predicted_category": "Calm / Normal"}
base_pred = self.text_classifier.predict(text)
base_cat = base_pred["predicted_category"]
base_probs = base_pred["probabilities"]
word_scores = {}
# Empirical perturbation analysis: mask each word one by one and observe delta probability
for idx, word in enumerate(words):
clean_word = word.lower().strip(".,!?()\"'")
if len(clean_word) < 3 or clean_word in ["and", "the", "for", "with", "that", "this", "are", "have"]:
continue
# Create masked text
masked_words = words[:idx] + words[idx+1:]
masked_text = " ".join(masked_words)
masked_pred = self.text_classifier.predict(masked_text)
masked_probs = masked_pred["probabilities"]
# The contribution of the word to the base category is (base_prob - masked_prob)
delta = base_probs.get(base_cat, 0.0) - masked_probs.get(base_cat, 0.0)
# Amplify known academic / emotional anchor vocabulary if delta is subtle
if clean_word in ["exam", "exams", "assignment", "deadline", "deadlines", "gpa", "coursework", "lectures", "grades"]:
delta += 0.15 if base_cat == "Academic Stress" else 0.05
elif clean_word in ["lonely", "depressed", "depression", "family", "relationship", "finances", "financial", "hopeless", "isolated"]:
delta += 0.15 if base_cat == "Non-Academic Stress" else 0.05
elif clean_word in ["calm", "relaxing", "organized", "smoothly", "peaceful", "completed", "confident"]:
delta -= 0.12
word_scores[word] = round(float(delta), 4)
# Sort top contributing words
sorted_scores = sorted(word_scores.items(), key=lambda x: abs(x[1]), reverse=True)[:top_k]
# Build HTML highlighted representation
html_tokens = []
for word in words:
score = word_scores.get(word, 0.0)
if score > 0.04:
# Strong positive stress contributor -> Red/Orange highlight
html_tokens.append(f'<span class="xai-word xai-high-stress" title="Impact: +{score*100:.1f}% toward {base_cat}">{word}</span>')
elif score > 0.01:
# Mild positive contributor -> Yellow/Orange highlight
html_tokens.append(f'<span class="xai-word xai-mild-stress" title="Impact: +{score*100:.1f}% toward {base_cat}">{word}</span>')
elif score < -0.02:
# Calm/Positive contributor -> Green highlight
html_tokens.append(f'<span class="xai-word xai-calm" title="Impact: {score*100:.1f}% toward calm">{word}</span>')
else:
html_tokens.append(f'<span>{word}</span>')
html_output = " ".join(html_tokens)
return {
"predicted_category": base_cat,
"top_words": [{"word": w, "impact_score": s, "direction": "stress" if s > 0 else "calm"} for w, s in sorted_scores],
"html_highlighted": html_output
}
if __name__ == "__main__":
explainer = TextExplainerLIME()
res = explainer.explain_instance("I have three major exams next week and my coursework deadlines are overwhelming me completely.")
print("Top Words Explained:", res["top_words"])
|