File size: 6,772 Bytes
cbb1b1a
 
 
 
 
 
4c85df9
cbb1b1a
 
4c85df9
 
 
cbb1b1a
 
 
4c85df9
 
 
cbb1b1a
4c85df9
 
 
 
 
 
cbb1b1a
4c85df9
 
 
cbb1b1a
 
 
4c85df9
 
 
 
 
 
 
 
 
 
 
 
 
 
cbb1b1a
 
 
 
 
 
 
 
 
 
 
4c85df9
 
 
 
cbb1b1a
4c85df9
 
 
 
 
 
 
cbb1b1a
4c85df9
 
 
cbb1b1a
4c85df9
 
 
1f54b5c
 
 
 
 
4c85df9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cbb1b1a
4c85df9
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
"""
EvidenceNER model definition.

Architecture: distilbert-base-uncased with token classification head.
Task:         Named Entity Recognition on redacted complaint text.
Entity types: ORG | AMOUNT | DATE | REF_ID | ACCOUNT | PERSON
Input:        Redacted complaint text (str, max 512 tokens after tokenisation).
Output:       List of Entity(text, label, start, end, confidence).

BIO scheme: O + B-/I- prefix for each of the 6 entity types β†’ 13 labels total.

Note: PERSON entities surviving Presidio redaction are role-references
      (e.g. "customer care executive"), not personal names.
"""

from __future__ import annotations

import logging
from dataclasses import dataclass
from typing import Optional

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

logger = logging.getLogger(__name__)

# ---------------------------------------------------------------------------
# Label constants β€” shared by model.py, train.py, predict.py
# ---------------------------------------------------------------------------

NER_LABELS = ["ORG", "AMOUNT", "DATE", "REF_ID", "ACCOUNT", "PERSON"]

# O first, then B-/I- pairs in the same order as NER_LABELS β†’ 13 labels
BIO_LABELS: list[str] = ["O"] + [
    f"{bio}-{label}" for label in NER_LABELS for bio in ("B", "I")
]

LABEL2ID: dict[str, int] = {label: i for i, label in enumerate(BIO_LABELS)}
ID2LABEL: dict[int, str] = {i: label for label, i in LABEL2ID.items()}

NUM_LABELS = len(BIO_LABELS)  # 13


# ---------------------------------------------------------------------------
# Public output type
# ---------------------------------------------------------------------------

@dataclass
class Entity:
    """A single recognised entity span."""
    text: str
    label: str
    start: int
    end: int
    confidence: float


# ---------------------------------------------------------------------------
# EvidenceNER
# ---------------------------------------------------------------------------

class EvidenceNER:
    """
    DistilBERT token classifier for complaint evidence extraction.

    Loads a fine-tuned checkpoint produced by train.py.  Uses the tokenizer's
    offset_mapping to convert subword-level BIO predictions back to
    character-level spans without any secondary tokenisation step.
    """

    BASE_MODEL = "distilbert-base-uncased"

    def __init__(self, model_dir: str) -> None:
        """Load a fine-tuned NER checkpoint from *model_dir*."""
        self.tokenizer = AutoTokenizer.from_pretrained(model_dir)
        self.model = AutoModelForTokenClassification.from_pretrained(model_dir)
        self.model.eval()
        self._device = torch.device(
            "cuda" if torch.cuda.is_available()
            else "mps" if torch.backends.mps.is_available()
            else "cpu"
        )
        self.model.to(self._device)
        logger.info("EvidenceNER loaded from %s on %s", model_dir, self._device)

    # ------------------------------------------------------------------
    # Public API
    # ------------------------------------------------------------------

    def extract(self, text: str) -> list[Entity]:
        """
        Extract entity spans from *text* and return a list of Entity objects.

        Spans are character-level (start/end index into the original string).
        Returns [] for empty or whitespace-only input.
        """
        if not text or not text.strip():
            return []

        encoding = self.tokenizer(
            text,
            return_tensors="pt",
            truncation=True,
            max_length=512,
            return_offsets_mapping=True,
        )

        # offset_mapping is not a model input β€” pop before forward pass
        offset_mapping: list[tuple[int, int]] = (
            encoding.pop("offset_mapping")[0].tolist()
        )

        model_inputs = {k: v.to(self._device) for k, v in encoding.items()}

        with torch.no_grad():
            logits = self.model(**model_inputs).logits[0]  # (seq_len, num_labels)

        probs = torch.softmax(logits, dim=-1).cpu()
        pred_ids: list[int] = probs.argmax(dim=-1).tolist()
        conf_scores: list[float] = probs.max(dim=-1).values.tolist()

        return self._aggregate_spans(text, offset_mapping, pred_ids, conf_scores)

    # ------------------------------------------------------------------
    # Span aggregation
    # ------------------------------------------------------------------

    def _aggregate_spans(
        self,
        text: str,
        offset_mapping: list[tuple[int, int]],
        pred_ids: list[int],
        conf_scores: list[float],
    ) -> list[Entity]:
        """
        Convert per-subtoken BIO predictions into character-level Entity spans.

        Special tokens ([CLS], [SEP]) have offset (0, 0) β€” i.e. start == end β€”
        and are skipped.  An I- tag that does not continue the current B- entity
        type is treated as O (broken sequence).
        """
        entities: list[Entity] = []
        current: Optional[dict] = None
        current_confs: list[float] = []

        def _flush() -> None:
            if current is not None:
                entities.append(Entity(
                    text=current["text"],
                    label=current["label"],
                    start=current["start"],
                    end=current["end"],
                    confidence=sum(current_confs) / len(current_confs),
                ))

        for (start, end), label_id, conf in zip(offset_mapping, pred_ids, conf_scores):
            # Special tokens have zero-length offset spans
            if start == end:
                _flush()
                current = None
                current_confs = []
                continue

            label = ID2LABEL[label_id]

            if label.startswith("B-"):
                _flush()
                entity_type = label[2:]
                current = {
                    "text": text[start:end],
                    "label": entity_type,
                    "start": start,
                    "end": end,
                }
                current_confs = [conf]

            elif (
                label.startswith("I-")
                and current is not None
                and label[2:] == current["label"]
            ):
                # Extend the current span (including any whitespace between subwords)
                current["text"] = text[current["start"]: end]
                current["end"] = end
                current_confs.append(conf)

            else:  # O or mismatched I- β†’ close current span
                _flush()
                current = None
                current_confs = []

        # Flush any span still open at end of sequence
        _flush()

        return entities