RansMairoi/khasi-ner

Khasi Named Entity Recognition Corpus

A named-entity-annotated corpus for the Khasi language, an Austroasiatic language spoken by approximately 1.6 million people, primarily in the state of Meghalaya in north-east India. To the best of the authors' knowledge, this is the first publicly described NER resource for Khasi.

The corpus was constructed as part of the PhD thesis "Machine Learning based Named Entity Recognition for Khasi Language" by Ransly Hoojon (North-Eastern Hill University, 2026) and underpins the empirical results reported in [Hoojon, 2025] and [Hoojon, Nath & Lyngdoh, 2026].

Dataset overview

Split	Documents	Tokens	Entity annotations
train	10,756	2,435,961	119,935
validation (test)	2,675	644,007	30,827
Total	13,431	3,079,968	150,762

(The entity counts in this release differ very slightly -- by 3 on the train split and 61 on the validation split -- from the figures reported in [Hoojon, Nath & Lyngdoh, 2026]; the discrepancy reflects minor cleaning carried out after publication and is well under 0.2% on either split.)

The 19 entity types extend the 18-class OntoNotes / spaCy core English NER tagset [Pradhan et al., 2013] with one Khasi-motivated addition, DAY, for Khasi day-of-the-week expressions (e.g. sngi u Blei 'Sunday', sngi thohdieng 'Friday').

Per-class distribution on the validation split

Tag	Count		Tag	Count
GPE	9,436		TIME	388
PERSON	6,046		DAY	303
ORG	3,675		PRODUCT	239
NORP	3,195		EVENT	191
LOC	2,134		QUANTITY	171
DATE	1,849		PERCENT	153
CARDINAL	1,631		FAC	126
MONEY	623		ORDINAL	82
LAW	521		LANGUAGE	48
			WORK_OF_ART	15

The distribution is sharply skewed: the most frequent class (GPE) has 9,436 mentions on the validation split, the least frequent (WORK_OF_ART) has 15. Aggregate metrics on this distribution should be reported as both micro- and macro-averaged F1 so that rare-class behaviour is visible.

Files

Path	Format	Purpose
data/train.jsonl	JSON-lines, one record per Doc with character-offset spans	Primary, framework-agnostic
data/dev.jsonl	same	Validation / test split
data/train.iob	CoNLL-style IOB2 TSV (token + tab + tag; blank line between Docs)	Drop-in for AllenNLP, Flair, transformers token-classification
data/dev.iob	same	Validation / test split
data/train.spacy	spaCy 3.x DocBin binary	Reproducibility supplement for spaCy users
data/dev.spacy	same	Validation / test split
data/stats.json	Per-split doc/token/entity counts	Cross-check

The three formats encode the same gold annotations. Pick whichever matches your training pipeline; reproducibility of the thesis results requires the .spacy files only.

JSONL record shape

{
  "text": "Shillong: Ha ka sngi balang ki bor district jong ka East Jaintia Hills District ki la pynpra noh ia ki artylli ki karkhana...",
  "spans": [
    {"start": 0,  "end": 8,  "label": "GPE"},
    {"start": 13, "end": 25, "label": "DAY"},
    {"start": 52, "end": 80, "label": "LOC"}
  ]
}

IOB2 block shape

Shillong	B-GPE
:	O
Ha	O
ka	O
sngi	B-DAY
balang	I-DAY
...
(blank line marks Doc boundary)

Loading

With Hugging Face datasets

from datasets import load_dataset

ds = load_dataset("YOUR_HANDLE/khasi-ner")
print(ds)
print(ds["train"][0])

With spaCy

import spacy
from spacy.tokens import DocBin

nlp = spacy.blank("en")
docs = list(DocBin().from_disk("data/train.spacy").get_docs(nlp.vocab))
print(len(docs), "documents")
for ent in docs[0].ents:
    print(ent.text, ent.label_, ent.start_char, ent.end_char)

Reading IOB2 with plain Python

def read_iob(path):
    sents, sent = [], []
    for line in open(path, encoding="utf-8"):
        line = line.rstrip("\n")
        if not line:
            if sent:
                sents.append(sent)
                sent = []
        else:
            tok, tag = line.split("\t")
            sent.append((tok, tag))
    if sent:
        sents.append(sent)
    return sents

train = read_iob("data/train.iob")
print(len(train), "sentences")
print(train[0][:6])

Tagset

Tag	Description	Khasi example
PERSON	People, including fictional characters	U Soso Tham
NORP	Nationalities or religious / political groups	Khasi, INC
FAC	Buildings, airports, highways, bridges	Logan Airport
ORG	Companies, agencies, institutions	KHADC
GPE	Countries, cities, states	Shillong, Meghalaya
LOC	Non-GPE locations (rivers, ranges, etc.)	Umngot
PRODUCT	Vehicles, foods, devices (non-services)	iPhone
EVENT	Battles, wars, sports events	FIFA World Cup
WORK_OF_ART	Titles of books, songs, paintings	Ka Akor Kaba Tam
LAW	Named documents made into laws	Roe v. Wade
LANGUAGE	Any named language	English, Khasi
DATE	Absolute or relative dates or periods	11 tarik naitung 2023
TIME	Times smaller than a day	12.00 baje
PERCENT	Percentages, including the % sign	80%
MONEY	Monetary amounts including currency	T.5 klur, T.2.5 lak
QUANTITY	Measurements of weight or distance	50 kg
ORDINAL	Ordered numbers	ba ar 'second'
CARDINAL	Numerals	saw ngut, saw tylli
DAY	Day of the week (Khasi-specific addition)	sngi u Blei 'Sunday'

Source

Raw text was collected from Khasi-language newspapers (primarily urupang.com and syllad.com) and Khasi-language books, and then annotated in Prodigy (Explosion AI) by native Khasi speakers under the supervision of a senior linguist. A hand-crafted pattern file of approximately 7,662 entity entries was used by Prodigy's ner.manual recipe to surface candidate spans for the annotator to confirm or reject; final entity decisions remained with the native-speaker annotator. Full construction details are in Chapter 2 of the accompanying thesis.

Known limitations

1. No quantitative inter-annotator agreement (IAA) figure is available in this release. The annotation budget for the project supported one annotator per sentence rather than the redundant double-annotation that an IAA computation requires, and the Prodigy installation was run in single-user mode so the database does not separate work by individual annotator. A Cohen's κ on a 100-sentence double-annotated sub-sample by a second native Khasi speaker is planned for the next revision of this dataset and will be reported in a CHANGELOG.
2. Topical bias toward contemporary newspaper prose. Religious, literary, scientific and historical Khasi texts are under-represented, and the resulting corpus is therefore stronger on PERSON and GPE mentions typical of news writing than on WORK_OF_ART or LAW mentions that would be more frequent in literary or legal text. The class-imbalance pattern above is partly a consequence of this topical bias rather than an absolute scarcity of those entities in Khasi in general.
3. Flat (non-overlapping) entity annotation only. Nested entities -- for example U Soso Tham (PERSON) inside U Soso Tham Auditorium (FAC) -- are tagged at the outermost level only. Nested entity recognition is identified as future work in the accompanying thesis.
4. Diacritic variability. Khasi orthography uses two diacritic characters, ñ and ï, which informal digital writing frequently replaces with plain n and i. Both spellings can appear in this corpus; downstream pipelines should normalise as appropriate for their task.

License

CC-BY-SA-4.0. You are free to share and adapt the corpus, including for commercial use, provided you give appropriate credit and distribute any derivatives under the same licence. See https://creativecommons.org/licenses/by-sa/4.0/ for full terms.

Citation

If you use this corpus, please cite the dataset (DOI below) and at least one of the two underlying papers:

@dataset{hoojon2026khasiner,
  author    = {Ransly Hoojon and Amitabha Nath and Saralin A. Lyngdoh},
  title     = {{K}hasi Named Entity Recognition Corpus},
  publisher = {Zenodo},
  year      = {2026},
  version   = {1.0.0},
  doi       = {10.5281/zenodo.20474594},
  url       = {https://doi.org/10.5281/zenodo.20474594},
}

@phdthesis{hoojon2026khasi,
  author  = {Ransly Hoojon},
  title   = {Machine Learning based Named Entity Recognition for {K}hasi Language},
  school  = {North-Eastern Hill University},
  address = {Shillong, Meghalaya, India},
  year    = {2026},
}

@article{hoojon2025ensemble,
  author    = {Ransly Hoojon and Amitabha Nath and Saralin A. Lyngdoh},
  title     = {Evaluating Ensemble-Based Classifiers for {K}hasi Named Entity Recognition with Data Imbalance Constraints},
  journal   = {Science and Technology Journal},
  publisher = {Mizoram University},
  volume    = {14},
  year      = {2026},
}

@inproceedings{hoojon2025neiais,
  author    = {Ransly Hoojon},
  title     = {Optimizing Low-Resource {K}hasi {NER} through Transfer Learning},
  booktitle = {Proceedings of the International Conference on North East India AI Summit (NEIAIS-2025)},
  address   = {Manipur, India},
  year      = {2025},
}

Contact

For questions, corrections or to report issues, open an issue on the dataset repository or contact the author at the institutional email listed in the thesis.