File size: 2,760 Bytes
6769fe7 271ceda 6769fe7 271ceda 6769fe7 |
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 |
# TIMTQE Benchmark Dataset
This repository provides the **TIMTQE benchmark dataset**, designed for **translation quality estimation (QE) of text images**.
TIMTQE consists of two complementary components:
- **MLQE-PE** – a large-scale synthetic dataset derived from MLQE-PE, where source sentences are rendered into text images and paired with translation quality annotations.
- **HistMTQE** – a human-annotated subset of historical documents (English–Chinese and Russian–Chinese), reflecting real-world noisy and degraded text image conditions.
Together, these resources enable benchmarking multimodal, multilingual QE models under both **synthetic** and **historical** settings.
---
## 📁 Data Organization
The dataset is structured into three main directories:
### 1. MLQE-PE_jsonl
Contains JSONL files organized by prompt type:
- `normal/` – default prompt format (used in main experiments)
- `cot/` – chain-of-thought style prompts
- `multi-task/` – multi-task prompt format
Each subdirectory includes:
- `train.json`
- `dev.json`
- test files (e.g., `test_en-de.json`)
**Fields in each JSONL entry:**
| Field | Description |
|------------|-------------|
| `image` | Relative path to the text image file |
| `text` | Prompt-response conversation (depends on prompt template) |
| `task_type`| Task identifier (e.g., `llava_sft`) |
---
### 2. MLQE-PE_image
Contains text image files under different augmentation settings:
- `pngs/` – default (normal) images
- `pngs_bleed_through/` – images with bleed-through noise
- `pngs_skew/` – images with skew distortion
- ... (9 variants in total, as detailed in the paper)
Each subdirectory represents one type of augmentation applied to MLQE-PE.
---
### 3. HistMTQE
Contains historical document test sets with human-annotated quality scores:
- `HistMTQE_en-zh_test.tsv` – English → Chinese test set
- `HistMTQE_ru-zh_test.tsv` – Russian → Chinese test set
**TSV fields:**
| Field | Description |
|--------------|-------------|
| `index` | Sample index |
| `original` | Source sentence |
| `translation`| Machine translation output |
| `scores` | Raw annotation scores from multiple annotators |
| `mean` | Average score |
| `z_scores` | Standardized scores |
| `z_mean` | Average standardized score |
---
## 🔄 Switching Data Variants
- **Change prompt template**
Select JSONL files from:
- `MLQE-PE_jsonl/normal/`
- `MLQE-PE_jsonl/cot/`
- `MLQE-PE_jsonl/multi-task/`
- **Change image augmentation**
Update the `image` path in JSONL files to point to the desired subdirectory.
|