anote-ai/TuluLegalRAG

TuluLegal

TuluLegal is a small, native-speaker-authored resource for studying cross-lingual legal comprehension in Tulu, a zero-resource Dravidian language spoken by roughly two million people in coastal Karnataka, India. Tulu has no legal corpus of its own — court proceedings for Tulu speakers are conducted in Kannada or English. This dataset was built to test whether an LLM can classify a Tulu speaker's everyday legal complaint into a legal issue category by retrieving context from an existing Kannada legal corpus, without full translation.

It accompanies the paper "Cross Lingual Transfer in Tulu Legal Comprehension: Script-Dependent Improvement and RAG-Induced Knowledge Conflict" (EMNLP 2026 ORACLE Workshop).

Dataset Description

The repository contains four files, covering two roles in the pipeline: queries (what a Tulu speaker says) and retrieval corpus (the Kannada legal documents a system would search over to understand it), plus a priming glossary used to scaffold model comprehension.

File	Role	Size
tulu_legal_60_consolidated.jsonl	Query set — 60 Tulu legal situations, 4 scripts each	60 rows
tulu_legal_20_consolidated.jsonl	Subset of the above — the original 20 sentences from the Phase 1 pilot	20 rows
kannada_legal_corpus_list.jsonl	Retrieval corpus — 18 real Karnataka court documents	18 rows
tulu_priming_glossary_trilingual_v2.json	Vocabulary glossary used for linguistic priming	32 entries

Languages

• Tulu (tcy) — the query language, provided in four scripts: Latin (romanized), Kannada, Tamil, and Malayalam
• Kannada (kn) — the administrative/retrieval-corpus language
• English (en) — glosses for every Tulu sentence, provided as a reference/control, not as a translation crutch for the pipeline itself

Dataset Structure

tulu_legal_60_consolidated.jsonl / tulu_legal_20_consolidated.jsonl

60 (and a 20-sentence subset) everyday legal situations, authored by a native Tulu speaker in an informal spoken register, each mapped to a primary and (where relevant) secondary category from the Stanford Legal Design Lab's LIST taxonomy, an access-to-justice classification organized around the person's problem rather than a legal instrument's clause.

Fields:

• id (int) — sentence identifier, consistent across both files
• english_gloss (string) — plain-English gloss of the situation (reference only)
• list_category (string) — primary LIST category
• list_category_secondary (string) — secondary LIST category, or empty string
• tulu_romanized (string) — Tulu in Latin script
• tulu_kannada (string) — Tulu written in Kannada script (the natural script for Tulu, historically written using Kannada abugidas)
• tulu_tamil (string) — Tulu transliterated into Tamil script
• tulu_malayalam (string) — Tulu transliterated into Malayalam script

LIST category distribution (60-sentence set):

Category	Count
Consumer/Money/Debt	10
Work/Employment	8
Torts/Individuals	8
Courts/Legal-System	8
Housing	7
Land/Property	7
IP/Identity	6
Estates/Wills	6

kannada_legal_corpus_list.jsonl

18 real Karnataka court documents (orders, judgments, hearings), sourced from ecourtsindia.com, ranging from single-paragraph procedural orders to 25-page witness hearings. Several source documents used legacy non-Unicode fonts that produced mojibake on direct extraction; text was extracted and then corrected with AI-assisted proofreading against the original document images, rather than open-ended inference, and category labels were manually assigned by the dataset author.

Fields:

• id (string) — short document identifier
• doc_number (int) — document index
• case_name (string) — case name
• cnr (string) — Case Number Record (India's court case identifier)
• cuad_categories (list of strings) — CUAD-derived category labels
• list_categories (list of strings) — LIST taxonomy category labels
• text (string) — full extracted document text (Kannada, with some English passages as they appear in the original bilingual court filings)
• char_count, word_count (int) — length of text
• retrieval_text (string) — cleaned version of text used for embedding/retrieval (boilerplate reduced)
• retrieval_word_count (int) — length of retrieval_text

Total corpus size: 18 documents, ~25,400 words combined.

LIST category coverage:

Category	Docs
Courts/Legal-System	8
Consumer/Money/Debt	5
Land/Property	5
Torts/Individuals	5
Housing	4
IP/Identity	4
Estates/Wills	3
Work/Employment	2

(Documents may carry more than one category, so counts sum to more than 18.)

tulu_priming_glossary_trilingual_v2.json

32 Tulu vocabulary entries (19 "safe" general-purpose words, 13 "leaky" legal/domain terms) used to linguistically prime models before classification, each given in Kannada, Tamil, and Malayalam script plus a romanized form and English meaning.

Fields:

• kannada, tamil, malayalam (string) — the word in each script
• romanized (string) — Latin-script form
• meaning (string) — English gloss
• tier (string) — "safe" (general vocabulary/grammar) or "leaky" (legal-domain terms that overlap with test-sentence content, used to isolate the effect of domain-specific priming)

Dataset Creation

Source Data

The Tulu sentences were authored directly by a native Tulu speaker for this project — they are not translated or machine-generated. Kannada- and English-gloss frontier model output was spot-checked against native-speaker judgment during construction; early drafts of the Kannada-script forms required correction on some sentences, underscoring that even frontier multilingual models cannot currently be trusted to produce reliable Tulu without native speaker verification.

The Kannada legal documents are real, public Karnataka court filings, scraped from ecourtsindia.com and cleaned via OCR + AI-assisted correction referenced against the original document images.

Annotations

LIST category labels (and CUAD-derived labels, where present) were assigned manually by the dataset author with reference to the Stanford Legal Design Lab taxonomy.

Considerations for Using the Data

• Scale. This is a small, pilot-scale resource (60 query sentences, 18 corpus documents) intended for research on cross-lingual comprehension methodology, not as a production legal-classification training set. Findings from it should be read as indicative trends, not high-confidence statistical claims — see the accompanying paper's limitations section.
• Not legal advice. These documents and sentence categorizations are for NLP research purposes only and must not be used to provide legal advice or make legal determinations about real cases.
• Language representation. Tulu is written here in four scripts to study script-dependent model comprehension; none of these scripts is a definitive standard orthography for Tulu, which does not have one in wide use.
• Community voice. This dataset exists specifically to study — and ultimately reduce — the erasure of Tulu speakers' original language when their situations enter a Kannada/English-mediated legal system. Users building on this data are encouraged to keep that framing central, and to involve native speakers in any downstream annotation or deployment decisions.

Licensing

This dataset is released under CC BY 4.0. You are free to share and adapt it for any purpose, including commercially, provided you give appropriate credit.

Citation

If you use this dataset, please cite:

@inproceedings{shetty2026tululegal,
  title     = {Cross Lingual Transfer in Tulu Legal Comprehension: Script-Dependent Improvement and RAG-Induced Knowledge Conflict},
  author    = {Shetty, Sindhu},
  booktitle = {Under Review},
  year      = {2026}
}

Contact

For questions about this dataset, please open a discussion on this repository or contact the author, Sindhu Shetty (Anote AI Research).