repleeka/tagin-monolingual-corpus

Tagin Monolingual Corpus

A ~110,000-segment monolingual text pool for Tagin, an endangered, highly agglutinative Tani language spoken in Arunachal Pradesh, India. The corpus was built to support Masked Language Modeling (MLM) / Continual Pre-Training (Domain-Adaptive Pre-Training, DAPT) of Transformer backbones on Tagin, as described 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).

Tagin has practically no prior digital footprint, so this corpus was collected and processed manually or semi-manually over an extended period — it is not scraped from the web.

Composition

The pool is built by combining three source channels, each contributing a different flavor of Tagin usage:

Source	Segments	Notes
GinLish Corpus v0.1	103,275	Tagin-English parallel lexicon; only the Tagin side is kept. ~70,549 (68%) entries are multi-word phrases; 33,698 unique Tagin surface forms.
Tagin Monolingual Corpus (community + transcribed audio)	~1,190	Digitized community write-ups, bulletins, narrative pieces, and hand-transcribed audio (conversations, folk stories, elder interviews). Richest in natural, conversational sentence structure.
Tagin NER Corpus (labels stripped)	10,600	The raw token stream from the gold NER corpus, rejoined into sentences with all NER labels removed.

Combined, these sources give ~115K segments before cleaning, settling to ~110K after filtering empty/duplicate rows. Each row in this dataset keeps a source field identifying which of the three channels it came from.

Why three sources?

Relying on a single source (e.g., only religious text or a single speaker's dialect) would bias a model toward a narrow style of language. Pooling formal narrative, spoken/conversational, and literary-translated (Bible-translation-derived) registers gives a wider stylistic spread, which is what Continual Pre-Training / DAPT needs.

• Community archive text is native-speaker-authored (not translated), capturing natural collocations and idiomatic phrasing.
• Audio transcription captures spoken, non-formal sentence patterns not well represented in written sources, using the project's modified Latin orthography (w/v vowel convention, vowel-doubling for long sounds).
• Biblical texts (portions of the Bible translated into Tagin by past missionary/community translation efforts) are one of the only sources of long-form, grammatically well-formed, internally consistent Tagin prose available in semi-digitized form; verses are treated as sentence-level units. This same translation effort also produced the Tagin-English word pairs behind GinLish Corpus v0.1.

Dataset structure

Each example has:

• text (string): a Tagin sentence or phrase-level segment.
• source (string): one of tagin_monolingual_corpus, ginlish_corpus_v0.1, or tagin_ner_corpus_raw.

Splits: train (90%) / validation (5%) / test (5%), shuffled with seed=42 — mirroring the 90/5/5 split the paper's MLM / DAPT stage evaluates on (103,790 / 5,766 / 5,767 out of a 115,323-segment pool).

Intended use

Masked Language Modeling / Continual Pre-Training (domain adaptation) of Transformer encoders on Tagin, as a precursor to downstream fine-tuning (e.g., NER — see the companion Tagin NER Corpus dataset). Not intended as clean, deduplicated running prose end-to-end — the GinLish-derived rows are short lexicon phrases, not full sentences.

Limitations

• Text is lowercased throughout.
• No standardized orthography existed prior to this project; transcription choices for ambiguous sounds reflect one transcriber's/team's judgment calls rather than an established writing convention.
• The GinLish-derived portion mixes single words and short phrases with the full-sentence segments from the other two sources.

How to use

from datasets import load_dataset

ds = load_dataset("repleeka/tagin-monolingual-corpus")
print(ds)
# DatasetDict({
#     train: Dataset({features: ['text', 'source'], num_rows: 103494})
#     validation: Dataset({features: ['text', 'source'], num_rows: 5750})
#     test: Dataset({features: ['text', 'source'], num_rows: 5750})
# })

example = ds["train"][0]
print(example["text"])
print(example["source"])

# Filter to a single source channel, e.g. only the natural-sentence portion:
community_only = ds["train"].filter(lambda ex: ex["source"] == "tagin_monolingual_corpus")

For MLM / Continual Pre-Training, tokenize the text column and pass it through a DataCollatorForLanguageModeling:

from transformers import AutoTokenizer, DataCollatorForLanguageModeling

tokenizer = AutoTokenizer.from_pretrained("MWirelabs/ne-bert")  # or any other backbone
tokenized = ds.map(
    lambda ex: tokenizer(ex["text"], truncation=True, max_length=128),
    batched=True,
    remove_columns=["text", "source"],
)
collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=True, mlm_probability=0.15)

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.