---
license: cc-by-nc-sa-4.0
language:
- zh
- nan
task_categories:
- translation
tags:
- machine-translation-evaluation
- error-span-detection
- MQM
- taiwanese-hokkien
- low-resource
- sinitic
pretty_name: Error Span Annotation for Taiwanese Hokkien
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/data.jsonl
dataset_info:
  features:
  - name: system
    dtype: string
  - name: id
    dtype: string
  - name: lp
    dtype: string
  - name: doc
    dtype: string
  - name: src
    dtype: string
  - name: mt
    dtype: string
  - name: ref
    dtype: string
  - name: spans
    list:
    - name: text
      dtype: string
    - name: start
      dtype: int64
    - name: end
      dtype: int64
    - name: severity
      dtype: string
    - name: category
      dtype: string
  - name: score100
    dtype: float64
  - name: score7
    dtype: float64
  - name: explanation
    dtype: string
  - name: rater
    dtype: string
  splits:
  - name: test
    num_examples: 154
---

# Error Span Annotation for Taiwanese Hokkien

*The Taiwanese Hokkien subset of the SiniticMTError benchmark (Liu et al., 2026).*

Human-annotated **machine-translation error-span** evaluation data for the
**Mandarin → Taiwanese Hokkien (Tâi-gí)** direction. Each instance contains a
Mandarin source sentence, a Taiwanese Hokkien machine translation, a reference
translation, and expert **error-span annotations** with severity labels and a
segment-level quality score.

- **Language pair:** Mandarin (`zh`) → Taiwanese Hokkien, Han script (`nan`)
- **MT system:** all machine translations (`mt`) were produced by `Taigi-Llama-2-Chat-7B`
- **Size:** 154 sentence pairs (single `test` split)
- **Annotation scheme:** MQM-based (Lommel et al., 2013) — error spans with a
  customized error-category set and two severity levels (`Minor` / `Major`);
  no `Critical` level is used

## Dataset structure

Each line in `data/data.jsonl` is one instance:

| Field | Type | Description |
|---|---|---|
| `system` | string | The MT system that produced `mt` (`Taigi-Llama-2-Chat-7B`) |
| `id` | string | Instance id |
| `lp` | string | Language pair code (`zh-han` = Mandarin → Taiwanese Hokkien, Han script) |
| `doc` | string | Source subset the sentence was drawn from (`moedict`) |
| `src` | string | Mandarin source sentence |
| `mt` | string | Taiwanese Hokkien machine translation (the hypothesis being evaluated) |
| `ref` | string | Reference Taiwanese Hokkien translation |
| `spans` | list | Error spans annotated on `mt` (see below); empty list means no error |
| `score7` | float | Segment-level quality score, 0–6 (SQM/DA style) |
| `score100` | float | Segment-level quality, 0–100; a linear rescale of `score7` (`score7 × 100/6`) |
| `explanation` | string | Free-text rationale for the annotated errors |
| `rater` | string | Annotator id (`rater1`, `rater2`) |

Each element of `spans` has:

| Field | Type | Description |
|---|---|---|
| `text` | string | The exact erroneous substring in `mt` |
| `start` | int | Character start offset in `mt` (0-based, inclusive) |
| `end` | int | Character end offset in `mt` (0-based, **exclusive**) |
| `severity` | string | `Minor`, `Major`, or `No-error` |
| `category` | string | MQM-style error type, e.g. `Accuracy/Mistranslation` |

Offsets index into `mt` at the **character** level; `mt[start:end] == text`.

### Example

```json
{
  "system": "Taigi-Llama-2-Chat-7B",
  "id": "0",
  "lp": "zh-han",
  "doc": "moedict",
  "src": "頭髮留那麼長還不剪，莫非是想把錢省下來買花生糖吃。",
  "mt": "頭毛留甲遐爾長毋去鉸，敢是欲儉錢來買塗豆麩食？",
  "ref": "頭毛留長長，儉錢食塗豆糖。",
  "spans": [
    {"text": "塗豆麩", "start": 18, "end": 21, "severity": "Major", "category": "Accuracy/Mistranslation"}
  ],
  "score100": 99.17,
  "score7": 5.95,
  "explanation": "1. 中文「花生糖」，台文是「塗豆糖」。",
  "rater": "rater1"
}
```

## Label sets

**Severity:** `Minor`, `Major`, `No-error` (no `Critical` level). In this subset:
Major 156, Minor 161, No-error 12.

**Error categories.** Annotators chose from an MQM-based typology (`Class/Subclass`):

- **Accuracy:** Mistranslation, Addition, Omission
- **Fluency:** Grammar, Spelling, Punctuation, Inconsistency, Register
- **Terminology:** Inappropriate, Inconsistent
- **Style:** Awkward
- **Locale:** Currency / Time / Name / Date / Address format
- **Purity:** Mandarin interference — Mandarin lexical items left untranslated in an
  otherwise Taiwanese Hokkien sentence (specific to this close language pair)

Not every category occurs in this 154-sentence subset. The data additionally uses
`Non-translated` and `No-error` markers, and a `.../Other` subcategory where an
error did not fit a listed subclass.

### Error-type distribution (306 error spans over 154 sentences)

| Top-level category | Major | Minor | Total |
|---|---:|---:|---:|
| Accuracy | 94 | 59 | 153 |
| Fluency | 11 | 48 | 59 |
| Purity | 22 | 14 | 36 |
| Locale | 2 | 29 | 31 |
| Terminology | 15 | 3 | 18 |
| Style | 1 | 4 | 5 |
| Other | 0 | 4 | 4 |
| **Total** | **145** | **161** | **306** |


## Collection & annotation

Mandarin source sentences and their Taiwanese Hokkien machine translations were
annotated by expert linguists (native Taiwanese Hokkien speakers). Annotators
first calibrated on shared examples to align on applying the MQM-based annotation
scheme to Taiwanese Hokkien, then annotated each sentence and resolved disagreements through
discussion to reach consensus. Because the two annotators were not independent,
conventional inter-annotator agreement is not reported.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("350016z/ErrorSpanAnnotation-for-Taiwanese-Hokkien", split="test")
print(ds[0]["src"], ds[0]["mt"])
for s in ds[0]["spans"]:
    print(s["text"], s["severity"], s["category"])
```

## Licensing & attribution

Released under **CC BY-NC-SA 4.0** (attribution, non-commercial, share-alike). This
reflects the licenses of the underlying materials:

- **Machine translations (`mt`)** were generated by **Taigi-Llama-2-Chat-7B**
  ([Bohanlu](https://huggingface.co/Bohanlu/Taigi-Llama-2-7B)), released under
  **CC BY-NC-SA 4.0**. Per the underlying **Llama 2** license and Acceptable Use
  Policy, these outputs may **not** be used to train or improve other large
  language models (other than Llama 2 or its derivatives).
- **Source sentences (`src`)** are derived from the **教育部臺灣台語常用詞辭典
  (MoE Taiwanese Dictionary)**, © Ministry of Education, R.O.C., and are included
  here for **non-commercial research** with attribution.
- **Annotations** (`spans`, `score100`, `score7`, `explanation`) are contributed by
  the dataset authors and released under **CC BY-NC-SA 4.0**.

This dataset is intended for **non-commercial research use only**.

> This is guidance, not legal advice. Some MoE dictionary materials carry a
> no-derivatives (ND) clause; if in doubt, confirm the terms with the Ministry of
> Education and your institution before redistribution.

## Citation

```bibtex
@article{liu2026siniticmterror,
  title   = {SiniticMTError: A Machine Translation Dataset with Error Annotations for Sinitic Languages},
  author  = {Liu, Hannah and Min, Junghyun and Lee, En-Shiun Annie and Cheung, Ethan Yue Heng and Hung, Shou-Yi and Chan, Elsie and Qian, Shiyao and Liang, Runtong and Huynh, Kimlan and Yip, Wing Yu and Ng, York Hay and Yau, Tsz Fung and Lo, Ka Ieng Charlotte and Wu, You-Wei and Tsai, Rich