"""
This module demonstrates a Streamlit application for masking Personally Identifiable
Information (PII) in Hebrew text using the GolemPII-v1 model.
"""
import time
from typing import List, Dict, Tuple
import streamlit as st
import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification
# Constants for model name and entity colors
MODEL_NAME = "CordwainerSmith/GolemPII-v1"
ENTITY_COLORS = {
"PHONE_NUM": "#FF9999",
"ID_NUM": "#99FF99",
"CC_NUM": "#9999FF",
"BANK_ACCOUNT_NUM": "#FFFF99",
"FIRST_NAME": "#FF99FF",
"LAST_NAME": "#99FFFF",
"CITY": "#FFB366",
"STREET": "#B366FF",
"POSTAL_CODE": "#66FFB3",
"EMAIL": "#66B3FF",
"DATE": "#FFB3B3",
"CC_PROVIDER": "#B3FFB3",
}
# Example sentences for demonstration
EXAMPLE_SENTENCES = [
"שם מלא: תלמה אריאלי מספר תעודת זהות: 61453324-8 תאריך לידה: 15/09/1983 כתובת: ארלוזורוב 22 פתח תקווה מיקוד 2731711 אימייל: mihailbenavi@ebox.co.il טלפון: 054-8884771 בפגישה זו נדונו פתרונות טכנולוגיים חדשניים לשיפור תהליכי עבודה. המשתתף יתבקש להציג מצגת בנושא בפגישה הבאה אשר שילם ב 5326-1003-5299-5478 מסטרקארד עם הוראת קבע ל 11-77-352300",
]
# Model details for display in the sidebar
MODEL_DETAILS = {
"name": "GolemPII-v1: Hebrew PII Detection Model",
"description": """
The GolemPII model
was specifically designed to identify and categorize various types of personally
identifiable information (PII) present in Hebrew text. Its core intended usage
revolves around enhancing privacy protection and facilitating the process of data
anonymization. This makes it a good candidate for applications and systems that
handle sensitive data, such as legal documents, medical records, or any text data
containing PII, where the automatic redaction or removal of such information is
essential for ensuring compliance with data privacy regulations and safeguarding
individuals' personal information. The model can be deployed on-premise with a
relatively small hardware footprint, making it suitable for organizations with
limited computing resources or those prioritizing local data processing.
The model was trained on the GolemGuard dataset, a Hebrew language dataset comprising over
115,000 examples of PII entities and containing both real and synthetically
generated text examples. This data represents various document types and
communication formats commonly found in Israeli professional and administrative
contexts. GolemGuard covers a wide range of document types and encompasses a
diverse array of PII entities, making it ideal for training and evaluating PII
detection models.
""",
"base_model": "xlm-roberta-base",
"training_data": "Custom Hebrew PII dataset",
"detected_pii_entities": [
"FIRST_NAME",
"LAST_NAME",
"STREET",
"CITY",
"PHONE_NUM",
"EMAIL",
"ID_NUM",
"BANK_ACCOUNT_NUM",
"CC_NUM",
"CC_PROVIDER",
"DATE",
"POSTAL_CODE",
],
}
class PIIMaskingModel:
"""
A class for masking PII in Hebrew text using the GolemPII-v1 model.
"""
def __init__(self, model_name: str):
"""
Initializes the PIIMaskingModel with the specified model name.
Args:
model_name: The name of the pre-trained model to use.
"""
self.model_name = model_name
hf_token = st.secrets["hf_token"]
self.tokenizer = AutoTokenizer.from_pretrained(model_name, token=hf_token)
self.model = AutoModelForTokenClassification.from_pretrained(
model_name, token=hf_token
)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.model.to(self.device)
self.model.eval()
def process_text(
self, text: str
) -> Tuple[str, float, str, List[str], List[str], List[Dict]]:
"""
Processes the input text and returns the masked text, processing time,
colored text, tokens, predicted labels, and privacy masks.
Args:
text: The input text to process.
Returns:
A tuple containing:
- masked_text: The text with PII masked.
- processing_time: The time taken to process the text.
- colored_text: The text with PII highlighted with colors.
- tokens: The tokens of the input text.
- predicted_labels: The predicted labels for each token.
- privacy_masks: A list of dictionaries containing information about
the masked PII entities.
"""
start_time = time.time()
tokenized_inputs = self.tokenizer(
text,
truncation=True,
padding=False,
return_tensors="pt",
return_offsets_mapping=True,
add_special_tokens=True,
)
input_ids = tokenized_inputs.input_ids.to(self.device)
attention_mask = tokenized_inputs.attention_mask.to(self.device)
offset_mapping = tokenized_inputs["offset_mapping"][0].tolist()
# Handle special tokens
offset_mapping[0] = None # token
offset_mapping[-1] = None # token
with torch.no_grad():
outputs = self.model(input_ids=input_ids, attention_mask=attention_mask)
predictions = outputs.logits.argmax(dim=-1).cpu().numpy()
predicted_labels = [
self.model.config.id2label[label_id] for label_id in predictions[0]
]
tokens = self.tokenizer.convert_ids_to_tokens(input_ids[0])
masked_text, colored_text, privacy_masks = self.mask_pii_in_sentence(
tokens, predicted_labels, text, offset_mapping
)
processing_time = time.time() - start_time
return (
masked_text,
processing_time,
colored_text,
tokens,
predicted_labels,
privacy_masks,
)
def _find_entity_span(
self,
i: int,
labels: List[str],
tokens: List[str],
offset_mapping: List[Tuple[int, int]],
) -> Tuple[int, str, int]:
"""
Finds the span of an entity starting at the given index.
Args:
i: The starting index of the entity.
labels: The list of labels for each token.
tokens: The list of tokens.
offset_mapping: The offset mapping for each token.
Returns:
A tuple containing:
- The index of the next token after the entity.
- The type of the entity.
- The end character offset of the entity.
"""
current_entity = labels[i][2:] if labels[i].startswith("B-") else labels[i][2:]
j = i + 1
last_valid_end = offset_mapping[i][1] if offset_mapping[i] else None
while j < len(tokens):
if offset_mapping[j] is None:
j += 1
continue
next_label = labels[j]
if next_label.startswith("B-") and tokens[j].startswith(" "):
break
if next_label.startswith("I-") and next_label[2:] != current_entity:
break
if next_label.startswith("I-") and next_label[2:] == current_entity:
last_valid_end = offset_mapping[j][1]
j += 1
elif next_label.startswith("B-") and not tokens[j].startswith(" "):
last_valid_end = offset_mapping[j][1]
j += 1
else:
break
return j, current_entity, last_valid_end
def mask_pii_in_sentence(
self,
tokens: List[str],
labels: List[str],
original_text: str,
offset_mapping: List[Tuple[int, int]],
) -> Tuple[str, str, List[Dict]]:
"""
Masks the PII entities in a sentence.
Args:
tokens: The list of tokens in the sentence.
labels: The list of labels for each token.
original_text: The original text of the sentence.
offset_mapping: The offset mapping for each token.
Returns:
A tuple containing:
- The masked text.
- The colored text.
- A list of dictionaries containing information about the masked
PII entities.
"""
privacy_masks = []
current_pos = 0
masked_text_parts = []
colored_text_parts = []
i = 0
while i < len(tokens):
if offset_mapping[i] is None:
i += 1
continue
current_label = labels[i]
if current_label.startswith(("B-", "I-")):
start_char = offset_mapping[i][0]
next_pos, entity_type, last_valid_end = self._find_entity_span(
i, labels, tokens, offset_mapping
)
if current_pos < start_char:
text_before = original_text[current_pos:start_char]
masked_text_parts.append(text_before)
colored_text_parts.append(text_before)
entity_value = original_text[start_char:last_valid_end]
mask = self._get_mask_for_entity(entity_type)
privacy_masks.append(
{
"label": entity_type,
"start": start_char,
"end": last_valid_end,
"value": entity_value,
"label_index": len(privacy_masks) + 1,
}
)
masked_text_parts.append(mask)
color = ENTITY_COLORS.get(entity_type, "#CCCCCC")
colored_text_parts.append(
f'{mask}'
)
current_pos = last_valid_end
i = next_pos
else:
if offset_mapping[i] is not None:
start_char = offset_mapping[i][0]
end_char = offset_mapping[i][1]
if current_pos < end_char:
text_chunk = original_text[current_pos:end_char]
masked_text_parts.append(text_chunk)
colored_text_parts.append(text_chunk)
current_pos = end_char
i += 1
if current_pos < len(original_text):
remaining_text = original_text[current_pos:]
masked_text_parts.append(remaining_text)
colored_text_parts.append(remaining_text)
return ("".join(masked_text_parts), "".join(colored_text_parts), privacy_masks)
def _get_mask_for_entity(self, entity_type: str) -> str:
"""
Returns the mask for a given entity type.
Args:
entity_type: The type of the entity.
Returns:
The mask for the entity type.
"""
return {
"PHONE_NUM": "[טלפון]",
"ID_NUM": "[ת.ז]",
"CC_NUM": "[כרטיס אשראי]",
"BANK_ACCOUNT_NUM": "[חשבון בנק]",
"FIRST_NAME": "[שם פרטי]",
"LAST_NAME": "[שם משפחה]",
"CITY": "[עיר]",
"STREET": "[רחוב]",
"POSTAL_CODE": "[מיקוד]",
"EMAIL": "[אימייל]",
"DATE": "[תאריך]",
"CC_PROVIDER": "[ספק כרטיס אשראי]",
"BANK": "[בנק]",
}.get(entity_type, f"[{entity_type}]")
def main():
"""
The main function for the Streamlit application.
"""
st.set_page_config(layout="wide")
st.title("🗿 GolemPII: Hebrew PII Masking Application 🗿")
st.markdown(
"""
""",
unsafe_allow_html=True,
)
# Sidebar with model details
st.sidebar.markdown(
f"""
{MODEL_DETAILS['description']}