Bangla Bagdhara
The largest computational resource for Bengali figurative language: 10,361 idioms annotated under a 19-field schema, released with the LREC 2026 benchmark.
What it contains
Why this dataset exists
Figurative language is one of the hardest things for large language models, and it is hardest in low-resource languages. Bengali is the seventh most spoken language in the world, yet before Bangla Bagdhara there was no large-scale, deeply annotated idiom corpus to train or test on.
Bangla Bagdhara is the largest computational resource for Bengali idioms: 10,361 expert-validated entries, released alongside the LREC 2026 benchmark in which none of 30 LLMs exceeded 50% accuracy, against 83.4% for human participants.
At a glance
| Idioms | 10,361 |
| Annotation schema | 19 fields per idiom |
| Tag assignments | 37,129 |
| Unique semantic tags | 9,919 |
| Format | JSON and JSONL (30.6 MB) |
| Licence | CC BY-SA 4.0 |
What each idiom carries
| Group | Fields |
|---|---|
| Identity | id, idiom, alternative forms |
| Meaning | literal meaning, figurative meaning in Bengali, figurative meaning in English, closest English equivalent |
| Usage | example sentences in Bengali and English, usage domain, frequency |
| Affect | sentiment, semantic tags |
| Context | historical, religious and cultural significance, geographic scope, history, notes |
Because every entry has both a gold figurative meaning and an expert literal gloss, the corpus works as a ready-made counterfactual test. Literal Lock-In uses exactly that to show models being pulled toward the literal reading.
Inside the corpus
Sentiment: 47.1% negative · 37.4% neutral · 15.5% positive
Cultural coverage:
- 99.7% of idioms are culturally significant
- 13.5% carry religious significance
- 6.3% carry historical significance
- Usage spans everyday conversation, literature, informal speech, media and education, across national, rural and urban registers.
How it was built
- Collected from authoritative sources spanning three centuries of Bengali: Bangla Academy dictionaries, classical and modern literature, regional folk collections, archival compilations and educational references.
- Drafted with AI assistance, every draft strictly human-reviewed.
- Validated by domain experts in Bangla linguistics through consensus, with every entry reviewed by at least two experts.
- Checked for 100% schema completeness on core fields.
Use it for
- Evaluating LLM figurative and cultural reasoning
- Bengali idiom detection and interpretation
- Cross-lingual figurative transfer and machine translation
- Sentiment-aware figurative modelling
- Bengali-focused RAG and educational tools
Cite
When Words Don't Mean What They Say: Figurative Understanding in Bengali Idioms, LREC 2026 · doi.org/10.63317/546w2cys6m6t
Dataset DOI: 10.34740/kaggle/dsv/14997176
Schema
19 fields
Unique tags
9,919
Expert review
2+ per entry
Official release
LREC 2026