Sentence-level relation extraction data for Tamil. Each item is one sentence from Tamil Wikipedia with two marked entity spans (subject and object) and one relation label between them.

Size and labels

Split	Items
train	3,407
validation	290
test	887

15 labels: NO_RELATION, LOC:LOCATED_IN, LOC:COUNTRY, PER:BIRTH_PLACE, PER:DEATH_PLACE, PER:RESIDENCE, PER:NATIONALITY, PER:EDUCATED_AT, PER:EMPLOYER, PER:MEMBER_OF, PER:PARENT, PER:SPOUSE, PER:SIBLING, ORG:HEADQUARTERS, ORG:FOUNDED_BY.

Entity types: PERSON, LOCATION, ORGANIZATION.

Format

One JSON object per line:

{
  "sentence": "...",
  "entities": [
    {"id": "E1", "text": "...", "type": "PERSON", "start": 0, "end": 9},
    {"id": "E2", "text": "...", "type": "LOCATION", "start": 46, "end": 57}
  ],
  "relation": {"subject": "E1", "label": "PER:NATIONALITY", "object": "E2"},
  "variety": "unknown",
  "source_id": "tawiki:<article title>#s<sentence index>",
  "meta": {...}
}

start/end are character offsets into sentence. Entity text is the surface form as it appears, so it often carries case suffixes (for example இந்தியாவைச் rather than இந்தியா).

meta fields: item_id, doc (source article), wd_subject / wd_object (Wikidata IDs), ds_label / ds_labels (the distant-supervision label the candidate was sampled with), match (whether each entity matched its Wikidata name exactly or in inflected form), annotator, decided_by, round, verified, notes, morph_analyzer, morph_pred (Stanza lemma/case predictions).

How it was built

Wikidata facts for the 14 relation types were matched against Tamil Wikipedia sentences (distant supervision), giving about 245k candidate pairs.
Candidates were sampled article-disjointly into two rounds (clean_v1, clean_v2).
Every item was labelled by Claude (Anthropic LLM) agents following annotation_guide.md, which allows correcting spans and entity types or relabelling as NO_RELATION. A double-labelled subset agreed at kappa 0.972 (clean_v1, 234 items) and 0.972 (clean_v2, 365 items). This agreement is between model runs, not between humans.
train = clean_v1 train + clean_v2 (deduplicated). validation and test = clean_v1 dev and test. Splits are article-disjoint.

Label quality: read this before using

All labels are LLM-produced (meta.annotator = "claude"), not human-annotated.
Human check: 300 test items were sampled for review by a Tamil-reading reviewer (data/human_review_test300.jsonl). All 300 were marked "agree". 25 of them (h001 to h025, mode: "review") were recorded one at a time. The other 275 (mode: "bulk_agree") were viewed earlier without a recorded verdict, and agreement with all of them was confirmed together afterwards. The remaining 587 test items and all train and validation items have no human check.
Known inconsistency: "X-ஐச் சேர்ந்த" (hails from) a city or state is labelled NO_RELATION in most items but PER:RESIDENCE in a few (about 5 of 23 in clean_v1).

Baseline numbers

Test micro-F1, entity-marker classifier, mean over 5 seeds, trained on full train: xlm-roberta-base 0.912, xlm-roberta-large 0.938.

License and attribution

Sentences come from Tamil Wikipedia (CC BY-SA 4.0); entity IDs and the facts used for sampling come from Wikidata (CC0). The dataset is released under CC BY-SA 4.0. Source article titles are in source_id / meta.doc for attribution.

Files

data/{train,dev,test}.jsonl: the splits.
data/human_review_test300.jsonl: human review verdicts for 300 test items (hid, item_id, the Claude label, human, mode).
annotation_guide.md: the labelling guide.

Downloads last month: 42

Total file size:

6.62 MB