Featured Datasets
Collection
5 items • Updated
Datasets:
Tasks:
Token Classification
Modalities:
Text
Sub-tasks:
part-of-speech
Languages:
Tagin
Size:
1K - 10K
DOI:
License:
Tagin POS Corpus v1.0
This repository contains the first Silver-Standard Part-of-Speech (POS) tagged corpus for the Tagin language (ISO 639-3: tgj), a low-resource Tani language spoken in Arunachal Pradesh, India and South-Tibet, China.
Dataset Summary
Due to the scarcity of digitally available text for Tagin, this corpus was constructed using 5,000 sentences curated from legacy dictionaries and grammar books. It serves as a benchmark for sequence labeling tasks in the Tani language family.
- Manuscript: A Silver-Standard POS Corpus and Baseline Models for Tagin, a Low-Resource Language (Under Review).
- Language: Tagin (
tgj) - Total Sentences: 5,000
- Total Tokens: 23,136
Corpus Statistics
| Statistic | Value |
|---|---|
| Total Sentences | 5,000 |
| Total Tokens | 23,136 |
| Vocabulary Size | 4,139 |
| Type-Token Ratio | 0.1789 |
| Average Sentence Length | 4.63 words |
Dataset Structure
Data Examples
The corpus is structured in a CSV format with two primary columns: the raw sentence and the corresponding tagged sequence.
| Sentence | Tagged_Sentence |
|---|---|
| tapa so aaba. | tapa/PROPN so/ADV aaba/VERB |
| aaba bv aato. | aaba/VERB bv/AUX aato/VERB |
| no aacho to. | no/PRON aacho/VERB to/PART |
| aadak lo jilaka. | aadak/VERB lo/ADP jilaka/VERB |
| aadak lvri bv aayo ke. | aadak/VERB lvri/NOUN bv/AUX aayo/VERB ke/PART |
| no so aadv aadv mabv ka. | no/PRON so/ADV aadv/VERB aadv/VERB mabv/AUX ka/PART |
| no aakor laku ka. | no/PRON aakor/VERB laku/AUX ka/PART |
Annotation Format
As shown in the snapshot, each token is followed by a forward slash (/) and its corresponding POS tag. This format is consistent throughout the 5,000 sentences of the GinLish MT project research.
POS Tagset
The corpus employs 21 hybrid tags compatible with UD v2, with specific morphosyntactic adaptations for Tagin:
Top 10 Most Frequent Tags:
- VERB (15.71%)
- NOUN (15.51%)
- PRON (14.14%)
- ADV (8.11%)
- AUX (7.92%)
- ADJ (6.98%)
- PART (6.44%)
- DET (6.16%)
- COP (5.36%)
- PROPN (4.51%)
Baseline Model Performance
The following classical sequence-labeling models were evaluated as baselines. CRF emerged as the most robust architecture for this dataset.
| Model Architecture | Word Acc (%) | Sent. Acc (%) | OOV Acc (%) | Macro-F1 (%) |
|---|---|---|---|---|
| CRF | 81.80 | 53.00 | 56.76 | 71.41 |
| Averaged Perceptron | 80.77 | 48.15 | 51.24 | 68.20 |
| MaxEnt | 80.05 | 47.90 | 50.11 | 67.54 |
| HMM | 77.18 | 40.80 | 39.62 | 63.12 |
Maintenance & Curation
- Curator: Tungon Dugi, Computer Science and Engineering Department, NIT Arunachal Pradesh.
- Project Name: GinLish MT.
Citation
If you use this dataset in your research, please cite:
@article{dugi2026tagin,
title={A Silver-Standard POS Corpus and Baseline Models for Tagin, a Low-Resource Language},
author={Dugi, Tungon and Sambyo, Koj},
journal={Computación y Sistemas (Under Review)},
year={2026}
}
- Downloads last month
- 9