sdk: static
title: TexTAR
emoji: 📚
license: mit
We introduce TexTAR, a multi-task, context-aware Transformer for Textual Attribute Recognition (TAR), capable of handling both positional cues (bold, italic) and visual cues (underline, strikeout) in noisy, multilingual document images.
MMTAD Dataset
MMTAD (Multilingual Multi-domain Textual Attribute Dataset) comprises 1,623 real-world document images—from legislative records and notices to textbooks and notary documents—captured under diverse lighting, layout, and noise conditions. It delivers 1,117,716 word-level annotations for two attribute groups:
T1: Bold,Italic,Bold & Italic
T2: Underline,Strikeout,Underline & Strikeout
Language & Domain Coverage
- English, Spanish, and six South Asian languages
- Distribution: 67.4 % Hindi, 8.2 % Telugu, 8.0 % Marathi, 5.9 % Punjabi, 5.4 % Bengali, 5.2 % Gujarati/Tamil/others
- 300–500 annotated words per image on average
To address class imbalance (e.g., fewer italic or strikeout samples), we apply context-aware augmentations:
- Shear transforms to generate additional italics
- Realistic, noisy underline and strikeout overlays
These augmentations preserve document context and mimic real-world distortions, ensuring a rich, balanced benchmark for textual attribute recognition.
More Information
For detailed documentation and resources, visit our website: TexTAR
Downloading the Dataset
from datasets import load_dataset
ds = load_dataset("textar/MMTAD")
print(ds)
Dataset contains
textar-testset: document imagestestset_labels.json: a JSON array or dict where each key/entry is an image filename and the value is its annotated attribute labels (bold, italic, underline, strikeout, etc. for each word)