src/scrub_entities.py

"""PERSON-only NER replace for allowlisted web sources.

FastPDN (ArkadiuszPawlak/fastpdn-ner-polish-pii, ONNX) tags person / street /
city / org. We replace only PERSON* and expand to the whole word so a
HerBERT hole cannot leave `[PII]ru[PII]`. Official and encyclopaedic sources
are default-deny — names there are the content.

Call after scrub_pii. Needs: pip install huggingface_hub tokenizers onnxruntime
"""
from __future__ import annotations

import json
import re
from pathlib import Path

from scrub_pii import PII_TAG

MODEL_ID = "ArkadiuszPawlak/fastpdn-ner-polish-pii"
PERSON_LABELS = frozenset({"PERSON", "PERSON_F", "PERSON_L"})
NER_SOURCES = frozenset({
    "european_hplt_v3_pl",
    "govpl",
    "samorzad_gov_pl",
})
_WORD = re.compile(r"[0-9A-Za-zÀ-ÿĄąĆćĘꣳŃńÓ󌜏źŻż'-]")

_NER = None


def source_allows_ner(source: str) -> bool:
    return source in NER_SOURCES


def _expand(text: str, start: int, end: int) -> tuple[int, int]:
    while start > 0 and _WORD.match(text[start - 1]):
        start -= 1
    while end < len(text) and _WORD.match(text[end]):
        end += 1
    return start, end


def apply_person_spans(text: str, spans: list[dict]) -> tuple[str, int]:
    """Replace PERSON* spans with [PII]. City/org/street spans are ignored."""
    kept: list[tuple[int, int]] = []
    for s in spans:
        if s.get("label") not in PERSON_LABELS:
            continue
        a, b = _expand(text, int(s["start"]), int(s["end"]))
        if a < b:
            kept.append((a, b))
    kept.sort()
    merged: list[tuple[int, int]] = []
    for a, b in kept:
        if merged and a <= merged[-1][1]:
            merged[-1] = (merged[-1][0], max(merged[-1][1], b))
        else:
            merged.append((a, b))
    out = text
    for a, b in reversed(merged):
        out = out[:a] + PII_TAG + out[b:]
    return out, len(merged)


def load_ner():
    import onnxruntime as ort
    from huggingface_hub import hf_hub_download
    from tokenizers import Tokenizer

    cfg = json.loads(Path(hf_hub_download(MODEL_ID, "config.json")).read_text())
    tok = Tokenizer.from_file(hf_hub_download(MODEL_ID, "tokenizer.json"))
    tok.enable_truncation(max_length=512)
    sess = ort.InferenceSession(
        hf_hub_download(MODEL_ID, "model_quantized.onnx"),
        providers=["CPUExecutionProvider"],
    )
    return {
        "sess": sess,
        "tok": tok,
        "id2label": {int(k): v for k, v in cfg["id2label"].items()},
    }


def _aggregate(text: str, labels: list[str], offsets, scores) -> list[dict]:
    spans = []
    cur = None
    for lab, (start, end), score in zip(labels, offsets, scores):
        if start == end or lab == "O" or "-" not in lab:
            if cur:
                spans.append(cur)
                cur = None
            continue
        prefix, typ = lab.split("-", 1)
        if cur and cur["label"] == typ and start <= cur["end"] + 1:
            cur["end"] = end
            cur["scores"].append(score)
        elif prefix == "B" or cur is None or cur["label"] != typ:
            if cur:
                spans.append(cur)
            cur = {"label": typ, "start": start, "end": end, "scores": [score]}
        else:
            cur["end"] = end
            cur["scores"].append(score)
    if cur:
        spans.append(cur)
    return [
        {
            "label": s["label"],
            "text": text[s["start"]:s["end"]],
            "score": round(sum(s["scores"]) / len(s["scores"]), 3),
            "start": s["start"],
            "end": s["end"],
        }
        for s in spans
    ]


def predict(ner, text: str) -> list[dict]:
    import numpy as np

    enc = ner["tok"].encode(text)
    ids = np.array([enc.ids], dtype=np.int64)
    mask = np.array([enc.attention_mask], dtype=np.int64)
    logits = ner["sess"].run(
        None,
        {
            "input_ids": ids,
            "attention_mask": mask,
            "token_type_ids": np.zeros_like(ids),
        },
    )[0][0]
    pred = logits.argmax(axis=-1)
    shift = logits - logits.max(axis=-1, keepdims=True)
    exp = np.exp(shift)
    prob = exp / exp.sum(axis=-1, keepdims=True)
    labels = [ner["id2label"][int(i)] for i in pred]
    scores = [float(prob[i, int(pred[i])]) for i in range(len(pred))]
    return _aggregate(text, labels, enc.offsets, scores)


def _ner():
    global _NER
    if _NER is None:
        _NER = load_ner()
    return _NER


def scrub_entities(
    text: str,
    source: str,
    spans: list[dict] | None = None,
) -> tuple[str, dict[str, int]]:
    """Return (text, {person: n}). No-op unless source is in NER_SOURCES."""
    counts = {"person": 0}
    if not text or not source_allows_ner(source):
        return text, counts
    if spans is None:
        spans = predict(_ner(), text)
    out, n = apply_person_spans(text, spans)
    counts["person"] = n
    return out, counts