File size: 6,473 Bytes
f4a4529
ef1515f
f4a4529
ea2a20f
f4a4529
 
 
ef1515f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea2a20f
 
 
ef1515f
 
 
ea2a20f
ef1515f
ea2a20f
ef1515f
 
ea2a20f
ef1515f
 
ea2a20f
ef1515f
 
ea2a20f
ef1515f
f4a4529
ea2a20f
 
 
 
2f22165
f4a4529
ea2a20f
 
 
 
f4a4529
ea2a20f
 
ef1515f
f4a4529
ea2a20f
 
 
 
 
 
 
f4a4529
ea2a20f
f4a4529
 
ea2a20f
f4a4529
 
 
 
 
 
 
 
ef1515f
f4a4529
ea2a20f
f4a4529
 
 
 
 
 
ef1515f
ea2a20f
ef1515f
 
 
f4a4529
 
 
 
 
 
 
 
2f22165
 
ef1515f
2f22165
 
 
ef1515f
2f22165
 
 
 
ef1515f
2f22165
 
 
ef1515f
f4a4529
ef1515f
 
ea2a20f
 
f4a4529
 
 
 
 
ea2a20f
f4a4529
ea2a20f
f4a4529
 
ef1515f
 
 
 
 
 
 
 
f4a4529
 
ef1515f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f4a4529
 
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
import json
import re
import argparse
from collections import Counter
import warnings
warnings.filterwarnings("ignore")


def normalize_prediction_text(pred_text):
    if pred_text is None:
        return "[EMPTY]"
    pred_text = pred_text.strip()
    return pred_text if pred_text else "[EMPTY]"


def exact_entity_match(ent_text, pred_text):
    if not ent_text or not pred_text:
        return False
    pattern = rf"(?<!\w){re.escape(ent_text)}(?!\w)"
    return re.search(pattern, pred_text, flags=re.IGNORECASE) is not None


def entity_leakage_rate(gold_records, predictions):
    leaked = 0
    total = 0
    per_type = Counter()
    leaked_per_type = Counter()

    for g, p in zip(gold_records, predictions):
        pred_text = normalize_prediction_text(p.get("anonymized_text", ""))
        for ent in g.get("entities", []):
            ent_text = ent.get("text", "")
            ent_type = ent.get("type", "UNKNOWN") or "UNKNOWN"
            total += 1
            per_type[ent_type] += 1
            if exact_entity_match(ent_text, pred_text):
                leaked += 1
                leaked_per_type[ent_type] += 1

    elr = leaked / total if total > 0 else 0.0
    return elr * 100, leaked, total, per_type, leaked_per_type

def get_capitalized_ngrams(text, n=3):
    """Extract n-grams where at least one token starts with uppercase."""
    tokens = text.split()
    ngrams = [tuple(tokens[i:i+n]) for i in range(len(tokens)-n+1)]
    return [ng for ng in ngrams if any(len(t) > 0 and t[0].isupper() for t in ng)]

def crr3(gold_records, predictions):
    """Capitalized 3-gram survival rate."""
    survived = 0
    total = 0
    
    for g, p in zip(gold_records, predictions):
        orig_3grams = get_capitalized_ngrams(g.get("original_text", ""), n=3)
        pred_text = normalize_prediction_text(p.get("anonymized_text", "")).lower()
        
        for ng in orig_3grams:
            total += 1
            if " ".join(ng).lower() in pred_text:
                survived += 1
                
    crr = survived / total if total > 0 else 0.0
    return crr * 100

def calculate_bertscore(gold_records, predictions, model_type="distilbert-base-uncased"):
    """
    BERTScore F1 computes the semantic similarity of the texts.
    Pinned model for reproducible benchmarking computations.
    """
    try:
        from bert_score import score
    except ImportError:
        print("`bert_score` not installed. Skipping. Install with: pip install bert_score")
        return None
        
    refs = [g.get("original_text", "") for g in gold_records]
    cands = [normalize_prediction_text(p.get("anonymized_text", "")) for p in predictions]
    
    P, R, F1 = score(cands, refs, lang="en", verbose=False, model_type=model_type)
    return F1.mean().item() * 100

def main():
    parser = argparse.ArgumentParser(description="SAHA-AL Benchmark Evaluator")
    parser.add_argument("--gold", type=str, default="data/test.jsonl", help="Path to gold dataset (e.g. test.jsonl)")
    parser.add_argument("--pred", type=str, required=True, help="Path to predictions JSONL")
    parser.add_argument("--bert-model", type=str, default="distilbert-base-uncased",
                        help="Model to use for BERTScore (e.g., microsoft/deberta-xlarge-mnli)")
    parser.add_argument("--print-types", action="store_true", help="Print per-entity-type ELR breakdown")
    parser.add_argument("--summary-file", type=str, default=None,
                        help="Optional JSON file to write evaluation results to")
    args = parser.parse_args()
    
    with open(args.gold, "r", encoding="utf-8") as f:
        gold_records = [json.loads(line) for line in f]
        
    with open(args.pred, "r", encoding="utf-8") as f:
        predictions = [json.loads(line) for line in f]
        
    unknown_entity_count = 0
    total_entity_count = 0

    for g, p in zip(gold_records, predictions):
        if g.get("id") != p.get("id"):
            raise ValueError(f"ID mismatch: {g.get('id')} vs {p.get('id')}")

        for ent in g.get("entities", []):
            total_entity_count += 1
            if ent.get("type", "UNKNOWN") == "UNKNOWN":
                unknown_entity_count += 1

    if total_entity_count > 0 and unknown_entity_count > 0:
        pct = (unknown_entity_count / total_entity_count) * 100
        print(f"[NOTE] {pct:.1f}% of entities in evaluation are typed as UNKNOWN.")

    print(f"Evaluating {len(predictions)} records...")

    elr, leaked, total_ents, per_type, leaked_per_type = entity_leakage_rate(gold_records, predictions)
    crr_3 = crr3(gold_records, predictions)
    bert_f1 = calculate_bertscore(gold_records, predictions, model_type=args.bert_model)
    
    print("\n" + "="*40)
    print(" SAHA-AL Benchmark Results")
    print("="*40)
    print(f" Entity Leakage Rate (ELR ↓): {elr:5.2f}% ({leaked}/{total_ents} leaked)")
    print(f" Contextual Re-ID (CRR-3 ↓):  {crr_3:5.2f}%")
    if bert_f1 is not None:
        print(f" BERTScore (F1 ↑):            {bert_f1:5.2f} (Model: {args.bert_model})")
    else:
        print(f" BERTScore (F1 ↑):            N/A")

    if args.print_types:
        print("\nPer-entity-type ELR:")
        for ent_type, count in per_type.most_common():
            leaked_count = leaked_per_type.get(ent_type, 0)
            elr_type = (leaked_count / count * 100) if count > 0 else 0.0
            print(f"  {ent_type:15} {elr_type:5.2f}% ({leaked_count}/{count})")

    print("="*40)

    if args.summary_file:
        summary = {
            "gold": args.gold,
            "predictions": args.pred,
            "records": len(predictions),
            "elr": round(elr, 2),
            "leaked": leaked,
            "total_entities": total_ents,
            "crr_3": round(crr_3, 2),
            "bert_f1": round(bert_f1, 2) if bert_f1 is not None else None,
            "bert_model": args.bert_model,
            "entity_types": {
                ent_type: {
                    "count": count,
                    "leaked": leaked_per_type.get(ent_type, 0),
                    "elr": round((leaked_per_type.get(ent_type, 0) / count * 100) if count > 0 else 0.0, 2),
                }
                for ent_type, count in per_type.items()
            },
        }
        with open(args.summary_file, "w", encoding="utf-8") as f:
            json.dump(summary, f, indent=2)
        print(f"Summary written to: {args.summary_file}")

if __name__ == "__main__":
    main()