File size: 4,401 Bytes
65a0ba8 0deb94b 16ae7ff 65a0ba8 | 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 | # MMTIT-Bench
**A Multilingual and Multi-Scenario Benchmark with Cognition–Perception–Reasoning Guided Text-Image Machine Translation**
<p align="center">
<a href="https://huggingface.co/datasets/VirtualLUO/MMTIT_Bench/blob/main/README_ZH.md">中文版</a> •
<a href="https://arxiv.org/abs/2603.23896">Paper</a> •
<a href="https://github.com/VirtualLUOUCAS/MMTIT_Bench">GitHub</a> •
<a href="https://huggingface.co/datasets/VirtualLUO/MMTIT_Bench">HuggingFace</a>
</p>
## Overview
**MMTIT-Bench** is a human-verified benchmark for end-to-end Text-Image Machine Translation (TIMT). It contains **1,400 images** spanning **14 non-English and non-Chinese languages** across diverse real-world scenarios, with bilingual (Chinese & English) translation annotations.
We also propose **CPR-Trans** (Cognition–Perception–Reasoning for Translation), a reasoning-oriented data paradigm that unifies scene cognition, text perception, and translation reasoning within a structured chain-of-thought framework.
<p align="center">
<img src="assets/overview.png" width="90%" alt="MMTIT-Bench Overview">
</p>
## Benchmark Statistics
| Item | Details |
|------|---------|
| Total Images | 1,400 |
| Languages | 14 (AR, DE, ES, FR, ID, IT, JA, KO, MS, PT, RU, TH, TR, VI) |
| Translation Directions | Other→Chinese, Other→English |
| Scenarios | Documents, Menus, Books, Attractions, Posters, Commodities, etc. |
| Annotation | Human-verified OCR + Bilingual translations |
## Data Format
### Directory Structure
```
MMTIT-Bench/
├── README.md
├── README_ZH.md
├── annotation.jsonl # Benchmark annotations
├── images.zip # Benchmark images
├── eval_comet_demo.py # COMET evaluation script
└── prediction_demo.jsonl # Example prediction file
```
### Annotation (`annotation.jsonl`)
Each line is a JSON object:
```json
{
"image_id": "Korea_Menu_20843.jpg",
"parsing_anno": "멜로우스트리트\n\n위치: 서울특별시 관악구...",
"translation_zh": "梅尔街\n\n位置：首尔特别市 冠岳区...",
"translation_en": "Mellow Street\n\nLocation: 1st Floor, 104 Gwanak-ro..."
}
```
| Field | Description |
|-------|-------------|
| `image_id` | Image filename, formatted as `{Language}_{Scenario}_{ID}.jpg` |
| `parsing_anno` | OCR text parsing annotation (source language) |
| `translation_zh` | Chinese translation |
| `translation_en` | English translation |
### Prediction File
Your prediction file should be a JSONL with the following fields:
```json
{"image_id": "Korea_Menu_20843.jpg", "pred": "Your model's translation output"}
```
## Evaluation
We use [COMET](https://github.com/Unbabel/COMET) (`Unbabel/wmt22-comet-da`) as the rule-based evaluation metric.
### Install
```bash
pip install unbabel-comet
```
### Run
```bash
# Other → Chinese
python eval_comet_demo.py \
--prediction your_prediction.jsonl \
--annotation annotation.jsonl \
--direction other2zh \
--batch_size 16 --gpus 0
# Other → English
python eval_comet_demo.py \
--prediction your_prediction.jsonl \
--annotation annotation.jsonl \
--direction other2en \
--batch_size 16 --gpus 1
```
### Arguments
| Argument | Default | Description |
|----------|---------|-------------|
| `--prediction` | *(required)* | Path to your prediction JSONL |
| `--annotation` | `annotation.jsonl` | Path to benchmark annotations |
| `--direction` | *(required)* | `other2zh` or `other2en` |
| `--batch_size` | `16` | Batch size for inference |
| `--gpus` | `0` | Number of GPUs (0 = CPU) |
| `--output` | `comet_results_{direction}.jsonl` | Output path for per-sample scores |
## Citation
```bibtex
@misc{li2026mmtitbench,
title={MMTIT-Bench: A Multilingual and Multi-Scenario Benchmark with Cognition-Perception-Reasoning Guided Text-Image Machine Translation},
author={Gengluo Li and Chengquan Zhang and Yupu Liang and Huawen Shen and Yaping Zhang and Pengyuan Lyu and Weinong Wang and Xingyu Wan and Gangyan Zeng and Han Hu and Can Ma and Yu Zhou},
year={2026},
journal={arXiv preprint arXiv:2603.23896},
url={https://arxiv.org/abs/2603.23896},
}
```
## License
This benchmark is released for **research purposes only**.
|