repleeka/tagin-ner-corpus

Tagin NER Corpus (CuNER Gold)

A 10,600-sentence, manually annotated, gold-standard Named Entity Recognition (NER) dataset for Tagin, an endangered, highly agglutinative Tani language spoken in Arunachal Pradesh, India. This is, to our knowledge, the first publicly available NER dataset for Tagin. Introduced in:

Tungon Dugi and Koj Sambyo. "Developing a Tagin NER Corpus and Benchmarking BERT-Based Models in a Low-Resource Setting." Sādhanā (Indian Academy of Sciences).

Annotation

A subset of sentences from the broader Tagin Monolingual Corpus pool was manually annotated token-by-token with BIO tags. Tagin has no capitalization cue to lean on (unlike English), so annotators relied purely on context and cultural/domain knowledge to decide entity boundaries. Sentences with ambiguous entity boundaries were cross-checked in a second pass. No formal inter-annotator agreement score was computed, since annotation was done by a small team rather than multiple fully independent annotators.

Tag scheme

Standard BIO scheme over five entity categories, for 11 distinct BIO labels total:

• PER — person names
• LOC — locations
• ORG — organizations
• DATE — dates/temporal expressions
• CULT — culturally specific entities (rituals, festivals, customs), a category introduced by this work. Many CULT names are built from ordinary descriptive words rather than dedicated proper nouns, so this category required the most annotator judgment.

O marks non-entity tokens.

Dataset structure

Each example has:

• id (string)
• tokens (list[string]): the sentence's tokens (surface word forms).
• ner_tags (list[ClassLabel]): one BIO tag per token, from ["O", "B-PER", "I-PER", "B-LOC", "I-LOC", "B-ORG", "I-ORG", "B-DATE", "I-DATE", "B-CULT", "I-CULT"].

A pos field exists in the raw source file but is currently empty (reserved for future work) and is dropped when building this dataset.

Splits: train (80%, 8,480 sentences) / validation (10%, 1,060 sentences) / test (10%, 1,060 sentences), seed=42. This reproduces the paper's exact NER fine-tuning split, including its two-step procedure: test is held out first (10% of the full corpus), then validation is held out from the remaining 90% (at a 1/9 proportion, equal to 10% of the original corpus), leaving 80% for training.

Statistics

• 10,600 sentences, 59,820 tokens (avg. ~5.6 tokens/sentence — short due to Tagin's agglutinative morphology, which packs meaning into single tokens via prefixing/suffixing).
• 10,182 unique word types.
• Sentence length is skewed: most sentences fall in the 3-10 token range, with a few outliers up to 237 tokens (full biblical verses/paragraphs kept as a single annotation unit).
• 11,413 total entity spans (B- tags) vs. 44,218 tokens tagged O (entities make up a little under 20% of all tokens).
• Entity span distribution is imbalanced: CULT 41%, PER 26%, LOC 16%, DATE 9%, ORG 8%. This reflects the source material (religious/community text discusses rituals and cultural practices far more than organizations), and makes ORG the hardest class for models to learn.
• Most entities are single-token: 8,133 of 11,413 spans are a single token; only 3,280 span two or more tokens.
• Vocabulary overlap with the GinLish Corpus v0.1 Tagin lexicon is partial: only 3,846 of the 10,182 unique word types (~38%) are shared.

Intended use

Training/evaluating token-classification (NER) models for Tagin, ideally after domain-adaptive pre-training on the companion Tagin Monolingual Corpus dataset (the two-stage transfer learning paradigm used in the paper). Because the label distribution is imbalanced (CULT-heavy), evaluation should report per-class metrics rather than relying on overall accuracy/F1 alone.

Limitations

• Text is lowercased throughout; there is no capitalization signal for entity detection.
• Class imbalance (CULT vs. ORG in particular).
• No formal inter-annotator agreement was computed.
• A handful of very long sentences (up to 237 tokens) come from biblical verses/paragraphs kept as single annotation units — account for this when choosing a max sequence length for fine-tuning.

How to use

from datasets import load_dataset

ds = load_dataset("repleeka/tagin-ner-corpus")
print(ds)
# DatasetDict({
#     train: Dataset({features: ['id', 'tokens', 'ner_tags'], num_rows: 8480})
#     validation: Dataset({features: ['id', 'tokens', 'ner_tags'], num_rows: 1060})
#     test: Dataset({features: ['id', 'tokens', 'ner_tags'], num_rows: 1060})
# })

example = ds["train"][0]
print(example["tokens"])
print(example["ner_tags"])  # integer class ids

# ner_tags is a Sequence(ClassLabel), so the string names travel with the
# dataset — no need to hardcode the label list yourself:
label_names = ds["train"].features["ner_tags"].feature.names
print(label_names)
print([label_names[i] for i in example["ner_tags"]])

For token-classification fine-tuning (e.g. with AutoModelForTokenClassification), build the label mappings directly from the feature:

label_names = ds["train"].features["ner_tags"].feature.names
id2label = dict(enumerate(label_names))
label2id = {v: k for k, v in id2label.items()}

Citation

@article{dugi_sambyo_tagin_ner,
  title   = {Developing a Tagin NER Corpus and Benchmarking BERT-Based Models in a Low-Resource Setting},
  author  = {Dugi, Tungon and Sambyo, Koj},
  journal = {S\={a}dhan\={a}},
  publisher = {Indian Academy of Sciences}
}

License

CC BY-NC-ND 4.0, not the repository's MIT code license — this dataset contains community-contributed linguistic material, not project source code.