File size: 9,427 Bytes
a05a9b8
 
 
 
 
 
 
 
 
 
dff5567
a05a9b8
dff5567
 
 
 
a05a9b8
 
 
 
dff5567
a05a9b8
 
 
dff5567
a05a9b8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dff5567
a05a9b8
 
 
 
 
 
 
 
 
 
 
 
dff5567
 
 
 
 
 
 
a05a9b8
 
 
dff5567
 
 
 
 
a05a9b8
 
 
dff5567
 
a05a9b8
dff5567
a05a9b8
 
 
 
 
 
 
 
dff5567
a05a9b8
dff5567
a05a9b8
 
dff5567
a05a9b8
 
 
 
dff5567
a05a9b8
 
 
 
 
 
 
 
dff5567
 
 
a05a9b8
 
 
 
 
 
 
 
 
dff5567
 
a05a9b8
 
dff5567
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a05a9b8
dff5567
 
 
 
 
a05a9b8
 
 
 
 
dff5567
 
 
 
 
 
 
a05a9b8
dff5567
 
 
 
 
 
 
 
 
 
 
 
 
a05a9b8
 
 
 
dff5567
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
29ae185
 
 
 
 
 
 
 
a05a9b8
 
 
dff5567
 
 
 
 
a05a9b8
 
 
 
 
29ae185
 
 
 
 
 
 
 
 
 
 
a05a9b8
29ae185
 
a05a9b8
 
dff5567
 
 
 
 
 
 
a05a9b8
 
 
 
dff5567
 
 
 
 
a05a9b8
 
 
 
 
 
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
"""
NERPA – Text anonymisation using the fine-tuned GLiNER2 model.

Usage:
    python anonymise.py "My name is John Smith, born 15/03/1990. Email: john@example.com"
    python anonymise.py --file input.txt
    python anonymise.py --file input.txt --output anonymised.txt
"""

import argparse
import logging
import sys
import warnings
from typing import Optional

warnings.filterwarnings("ignore", message=r".*incorrect regex pattern.*fix_mistral_regex.*")

import torch
from gliner2 import GLiNER2

logger = logging.getLogger(__name__)

# Entity types the model was fine-tuned to recognise, with descriptions
# that guide the bi-encoder towards better detection.
PII_ENTITIES: dict[str, str] = {
    "LOCATION": "Address, country, city, postcode, street, any other location",
    "AGE": "Age of a person",
    "DIGITAL_KEYS": "Digital keys, passwords, pins used to access anything like servers, banks, APIs, accounts etc",
    "BANK_ACCOUNT_DETAILS": "Bank account details such as number, IBAN, SWIFT, routing numbers etc",
    "CARD_DETAILS": "Debit or credit card details such as card number, CVV, expiration etc",
    "DATE_TIME": "Generic date and time",
    "DATE_OF_BIRTH": "Date of birth",
    "PERSONAL_ID_NUMBERS": "Common personal identification numbers such as passport numbers, driving licenses, taxpayer and insurance numbers",
    "TECHNICAL_ID_NUMBERS": "IP and MAC addresses, serial numbers and any other technical ID numbers",
    "EMAIL": "Email",
    "PERSON_NAME": "Person name",
    "BUSINESS_NAME": "Business name",
    "PHONE": "Any personal or other phone numbers",
    "URL": "Any short or full URL",
    "USERNAME": "Username",
    "VEHICLE_ID_NUMBERS": "Any vehicle numbers like license plates, vehicle identification numbers",
}

CONFIDENCE_THRESHOLD = 0.25
CHUNK_SIZE = 3000
CHUNK_OVERLAP = 100
BATCH_SIZE = 32


def load_model(model_path: str = ".") -> GLiNER2:
    """Load the NERPA model onto the best available device."""
    if torch.cuda.is_available():
        device = torch.device("cuda")
    elif torch.backends.mps.is_available():
        device = torch.device("mps")
    else:
        device = torch.device("cpu")

    model = GLiNER2.from_pretrained(model_path)
    try:
        model.to(device)
    except RuntimeError:
        logger.warning(
            "Failed to load model on %s, falling back to CPU.", device
        )
        model.to(torch.device("cpu"))
    return model


def chunk_text(
    text: str,
    chunk_size: int = CHUNK_SIZE,
    overlap: int = CHUNK_OVERLAP,
) -> tuple[list[str], list[int]]:
    """Split text into overlapping chunks, returning chunks and their start offsets."""
    if not text:
        return [], []
    chunks: list[str] = []
    starts: list[int] = []
    step = chunk_size - overlap
    for pos in range(0, len(text), step):
        chunks.append(text[pos : pos + chunk_size])
        starts.append(pos)
    return chunks, starts


def detect_entities(
    model: GLiNER2,
    text: str,
    entities: Optional[dict[str, str]] = None,
    threshold: float = CONFIDENCE_THRESHOLD,
) -> list[dict]:
    """
    Detect PII entities in text, returning a list of
    ``{"type": str, "start": int, "end": int, "score": float}`` dicts
    with character offsets into the original text.
    """
    entities = entities or PII_ENTITIES

    # Always detect both date types so the model can disambiguate.
    detect = dict(entities)
    if "DATE_TIME" in detect and "DATE_OF_BIRTH" not in detect:
        detect["DATE_OF_BIRTH"] = PII_ENTITIES["DATE_OF_BIRTH"]
    elif "DATE_OF_BIRTH" in detect and "DATE_TIME" not in detect:
        detect["DATE_TIME"] = PII_ENTITIES["DATE_TIME"]

    chunks, offsets = chunk_text(text)

    all_chunk_results: list[dict] = []
    for batch_start in range(0, len(chunks), BATCH_SIZE):
        batch = chunks[batch_start : batch_start + BATCH_SIZE]
        results = model.batch_extract_entities(
            batch,
            detect,
            include_confidence=True,
            include_spans=True,
            threshold=threshold,
        )
        all_chunk_results.extend(results)

    # Merge results across chunks: de-duplicate overlapping detections.
    seen: dict[tuple[int, int], dict] = {}
    for chunk_result, chunk_offset in zip(all_chunk_results, offsets):
        for label, occurrences in chunk_result["entities"].items():
            for occurrence in occurrences:
                start = occurrence["start"] + chunk_offset
                end = occurrence["end"] + chunk_offset
                position = (start, end)
                if (
                    position not in seen
                    or seen[position]["score"] < occurrence["confidence"]
                ):
                    seen[position] = {
                        "type": label,
                        "score": occurrence["confidence"],
                    }

    # Merge overlapping spans, keeping the highest-confidence label.
    # NOTE: when two spans overlap they are fused into one span and
    # assigned the label with the higher confidence score.
    items = sorted(
        [
            (start, end, info)
            for (start, end), info in seen.items()
            if info["type"] in entities
        ],
        key=lambda x: (x[0], x[1]),
    )
    if not items:
        return []

    merged: list[dict] = []
    current_start, current_end, current_info = items[0]
    for start, end, info in items[1:]:
        if start < current_end:  # overlapping
            current_end = max(current_end, end)
            if info["score"] > current_info["score"]:
                current_info = info
        else:
            merged.append({
                "type": current_info["type"],
                "start": current_start,
                "end": current_end,
                "score": current_info["score"],
            })
            current_start, current_end, current_info = start, end, info
    merged.append({
        "type": current_info["type"],
        "start": current_start,
        "end": current_end,
        "score": current_info["score"],
    })

    return merged


def anonymise(text: str, detected: list[dict]) -> str:
    """Replace detected entities with placeholders like ``[PERSON_NAME]``."""
    parts: list[str] = []
    prev_end = 0
    for entity in sorted(detected, key=lambda e: e["start"]):
        parts.append(text[prev_end : entity["start"]])
        parts.append(f'[{entity["type"]}]')
        prev_end = entity["end"]
    parts.append(text[prev_end:])
    return "".join(parts)


def main() -> None:
    parser = argparse.ArgumentParser(
        description="Anonymise PII in text using the NERPA model.",
    )
    parser.add_argument(
        "text", nargs="?", help="Text to anonymise (or use --file)",
    )
    parser.add_argument(
        "--file", "-f", help="Read text from a file instead",
    )
    parser.add_argument(
        "--output", "-o",
        help="Write anonymised text to file (default: stdout)",
    )
    parser.add_argument(
        "--model", "-m", default=".",
        help="Path to model directory (default: current dir)",
    )
    parser.add_argument(
        "--threshold", "-t", type=float, default=CONFIDENCE_THRESHOLD,
        help=f"Confidence threshold (default: {CONFIDENCE_THRESHOLD})",
    )
    parser.add_argument(
        "--show-entities", action="store_true",
        help="Print detected entities before anonymised text",
    )
    parser.add_argument(
        "--extra-entities", "-e", action="append", metavar="LABEL=DESCRIPTION",
        help=(
            "Additional custom entity types to detect alongside the built-in "
            "PII entities. Repeat for each type.  Format: LABEL=\"Description\".  "
            "Example: -e PRODUCT=\"Product name\" -e SKILL=\"Professional skill\""
        ),
    )
    args = parser.parse_args()

    if args.file:
        try:
            with open(args.file, encoding="utf-8") as f:
                text = f.read()
        except OSError as exc:
            sys.exit(f"Error reading {args.file}: {exc}")
    elif args.text:
        text = args.text
    else:
        parser.error("Provide text as an argument or use --file")

    extra: dict[str, str] = {}
    if args.extra_entities:
        for item in args.extra_entities:
            if "=" not in item:
                parser.error(
                    f"Invalid --extra-entities value '{item}'. "
                    "Expected format: LABEL=\"Description\""
                )
            label, description = item.split("=", 1)
            extra[label.strip()] = description.strip()

    model = load_model(args.model)
    all_entities = {**PII_ENTITIES, **extra} if extra else None
    detected = detect_entities(model, text, entities=all_entities, threshold=args.threshold)

    if args.show_entities:
        for entity in detected:
            span = text[entity["start"] : entity["end"]]
            logger.info(
                "  %-25s [%5d:%5d] (score=%.2f)  %r",
                entity["type"], entity["start"], entity["end"],
                entity["score"], span,
            )

    result = anonymise(text, detected)

    if args.output:
        try:
            with open(args.output, "w", encoding="utf-8") as f:
                f.write(result)
        except OSError as exc:
            sys.exit(f"Error writing {args.output}: {exc}")
    else:
        print(result)


if __name__ == "__main__":
    main()