Spaces:
Sleeping
Sleeping
feat: Add utility functions for text processing and model prediction
Browse files- utils/__init__.py +4 -0
- utils/model_utils.py +16 -0
- utils/text_processing.py +45 -0
utils/__init__.py
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# utils/__init__.py
|
2 |
+
|
3 |
+
from .text_processing import extract_text_from_pdf, split_into_clauses
|
4 |
+
from .model_utils import predict_unfairness
|
utils/model_utils.py
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
def predict_unfairness(text, model, tokenizer):
|
4 |
+
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)
|
5 |
+
|
6 |
+
model.eval()
|
7 |
+
with torch.no_grad():
|
8 |
+
outputs = model(**inputs)
|
9 |
+
|
10 |
+
probabilities = torch.softmax(outputs.logits, dim=-1).squeeze()
|
11 |
+
predicted_class = torch.argmax(probabilities).item()
|
12 |
+
|
13 |
+
label_mapping = {0: 'clearly_fair', 1: 'potentially_unfair', 2: 'clearly_unfair'}
|
14 |
+
predicted_label = label_mapping[predicted_class]
|
15 |
+
|
16 |
+
return predicted_label, probabilities.tolist()
|
utils/text_processing.py
ADDED
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import PyPDF2
|
2 |
+
import spacy
|
3 |
+
import re
|
4 |
+
|
5 |
+
nlp = spacy.load("en_core_web_sm")
|
6 |
+
|
7 |
+
def extract_text_from_pdf(pdf_file):
|
8 |
+
reader = PyPDF2.PdfReader(pdf_file)
|
9 |
+
text = ""
|
10 |
+
for page in reader.pages:
|
11 |
+
text += page.extract_text()
|
12 |
+
return text
|
13 |
+
|
14 |
+
def split_into_clauses(text):
|
15 |
+
# Preprocess the text
|
16 |
+
text = re.sub(r'\s+', ' ', text) # Remove extra whitespace
|
17 |
+
text = re.sub(r'\n+', '\n', text) # Remove extra newlines
|
18 |
+
|
19 |
+
# Use spaCy to parse the text
|
20 |
+
doc = nlp(text)
|
21 |
+
|
22 |
+
clauses = []
|
23 |
+
current_clause = []
|
24 |
+
|
25 |
+
for sent in doc.sents:
|
26 |
+
current_clause.append(sent.text)
|
27 |
+
|
28 |
+
# Check if this sentence ends a clause
|
29 |
+
if re.search(r'\d+\.|\([a-z]\)|\([iv]+\)', sent.text) or len(' '.join(current_clause)) > 200:
|
30 |
+
clauses.append(' '.join(current_clause))
|
31 |
+
current_clause = []
|
32 |
+
|
33 |
+
# Add any remaining text as the last clause
|
34 |
+
if current_clause:
|
35 |
+
clauses.append(' '.join(current_clause))
|
36 |
+
|
37 |
+
# Post-process clauses
|
38 |
+
cleaned_clauses = []
|
39 |
+
for clause in clauses:
|
40 |
+
# Remove leading/trailing whitespace and numbers
|
41 |
+
clause = re.sub(r'^\s*\d+\.?\s*', '', clause.strip())
|
42 |
+
if clause:
|
43 |
+
cleaned_clauses.append(clause)
|
44 |
+
|
45 |
+
return cleaned_clauses
|