File size: 5,327 Bytes
7168de4
 
 
 
 
 
4bfd31e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: mit
task_categories:
- question-answering
size_categories:
- 1K<n<10K
---

# BMIKE-53: Cross-Lingual In-Context Knowledge Editing Benchmark


[![Paper (ACL 2025)](https://img.shields.io/badge/Paper-ACL%202025-blue)](https://arxiv.org/pdf/2406.17764)  [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)

This repo presents the code and data for the ACL 2025 paper: "[*BMIKE-53: Investigating Cross-Lingual Knowledge Editing with In-Context Learning*](https://arxiv.org/pdf/2406.17764)".

## Overview
**BMIKE-53** is the first comprehensive multilingual benchmark for **Cross-Lingual In-Context Knowledge Editing (IKE)**. It covers 53 languages and unifies three widely used knowledge editing datasets (zsRE, CounterFact, WikiFactDiff) into a consistent, multilingual format. This resource enables systematic evaluation and analysis of knowledge editing capabilities in large language models (LLMs) across a broad spectrum of languages and knowledge scenarios.

## Table of Contents

- [Datasets](#datasets)
- [Supported Languages](#supported-languages)
- [Benchmark Structure](#benchmark-structure)
- [How to Use](#how-to-use)
- [Citation](#citation)
- [License](#license)

---

## Datasets

- **zsRE:** Zero-shot relation extraction, regular fact modifications.
- **CounterFact:** Counterfactual (fabricated) fact edits for testing knowledge locality.
- **WikiFactDiff:** Real-world, temporally dynamic factual updates derived from WikiData.

---

## Supported Languages

BMIKE-53 covers 53 languages, including:

`af`, `ar`, `az`, `be`, `bg`, `bn`, `ca`, `ceb`, `cs`, `cy`, `da`, `de`, `el`, `es`, `et`, `eu`, `fa`, `fi`, `fr`, `ga`, `gl`, `he`, `hi`, `hr`, `hu`, `hy`, `id`, `it`, `ja`, `ka`, `ko`, `la`, `lt`, `lv`, `ms`, `nl`, `pl`, `pt`, `ro`, `ru`, `sk`, `sl`, `sq`, `sr`, `sv`, `ta`, `th`, `tr`, `uk`, `ur`, `vi`, `zh`, `en`.

See [data/lang.json](data/lang.json) for details.

---

## Benchmark Structure

### Data Format

Each BMIKE-53 data sample is structured as a dictionary with entries for each language (e.g., `"en"` for English, `"de"` for German, etc.). Each language-specific entry contains:

- `case_id`: Unique identifier for the knowledge edit case.
- `subject`: The subject entity or person the knowledge edit is about.
- `src`: The original query (reliability type), directly matching the edited knowledge.
- `rephrase`: A paraphrased version of the original query, testing generalization.
- `old`: The original (pre-edit) answer to the query.
- `alt`: The updated (post-edit) answer reflecting the new knowledge.
- `loc`: A locality control query about an unrelated subject, testing if unrelated knowledge is preserved.
- `loc_ans`: The correct answer for the locality query.
- `port`: A portability query about a related (one-hop) aspect, testing the transfer of edited knowledge.
- `port_ans`: The correct answer for the portability query.

All values are language-specific translations, and the same structure is repeated for each supported language.

**Example:**
```json
{
  "en": {
    "case_id": 5,
    "subject": "Bhagwant Mann",
    "src": "What position did Bhagwant Mann hold?",
    "rephrase": "What role did Bhagwant Mann serve in?",
    "old": "Member of the 17th Lok Sabha",
    "alt": "Chief Minister of Punjab",
    "loc": "Who was Dipsinh Shankarsinh Rathod?",
    "loc_ans": "Member of the 17th Lok Sabha",
    "port": "Who is the current officeholder of the position previously held by Bhagwant Mann?",
    "port_ans": "Amarinder Singh"
  },
  "de": {
    "case_id": 5,
    "subject": "Bhagwant Mann",
    "src": "Welche Position hatte Bhagwant Mann inne?",
    "rephrase": "Welche Rolle hat Bhagwant Mann ausgeübt?",
    "old": "Mitglied des 17. Lok Sabha",
    "alt": "Chief Minister von Punjab",
    "loc": "Wer war Dipsinh Shankarsinh Rathod?",
    "loc_ans": "Mitglied des 17. Lok Sabha",
    "port": "Wer ist der aktuelle Amtsinhaber der Position, die zuvor von Bhagwant Mann gehalten wurde?",
    "port_ans": "Amarinder Singh"
  }
}
```

### Files

- Benchmark data is organized under the `/data/BMIKE53/` directory, with three subfolders corresponding to each dataset:  
  - `/CounterFact/`
  - `/WikiFactDiff/`
  - `/zsRE/`
- Each dataset folder contains test files for all 53 languages, named using the pattern:  
  - `zsre_test_{lang}.json` (e.g., `zsre_test_en.json`, `zsre_test_de.json`, etc.)
  - `counterfact_test_{lang}.json`
  - `wikifactdiff_test_{lang}.json`
- Each file contains the test set for one language in JSON format, following the unified BMIKE-53 data structure.
- Example:
  ```
  data/BMIKE53/zsRE/zsre_test_en.json
  data/BMIKE53/zsRE/zsre_test_de.json
  data/BMIKE53/CounterFact/counterfact_test_en.json
  data/BMIKE53/WikiFactDiff/wikifactdiff_test_en.json
  ```

---

## Citation

If you use BMIKE-53, please cite:

```bibtex
@inproceedings{nie2025bmike53,
  title={BMIKE-53: Investigating Cross-Lingual Knowledge Editing with In-Context Learning},
  author={Nie, Ercong and Shao, Bo and Wang, Mingyang and Ding, Zifeng and Schmid, Helmut and Sch{\"u}tze, Hinrich},
  booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)},
  year={2025}
}
```

---

## License

This project is licensed under the [MIT License](LICENSE).