README.md

---
license: mit
task_categories:
- text-generation
language:
- ne
tags:
- nepali
- morphology
- root-affix
- devanagari
- low-resource
pretty_name: 'Nepali Root-Affix Dictionary '
size_categories:
- 10K<n<100K
---
# Nepali Root-Affix Dictionary

A word-level morphological segmentation dataset for Nepali, mapping surface
words to their root and affix components (e.g. `अँकाइनु` → `अँका` + `इनु`).

## Dataset Details

### Dataset Description

This dataset was built by scraping the **10th edition of the Nepali
dictionary**, parsing the scraped entries into structured word/root pairs,
and then applying a **heuristic rule engine** to extend root-affix coverage
to surface forms not explicitly listed as dictionary headwords (e.g.
inflected or derived forms of a root that *is* in the dictionary). Each row
records the raw surface word, its root, the affix (if any), the
`+`-delimited segmented string, and which process produced that row.

- **Language:** Nepali (ne)
- **Curated by:** [w4ashabii](https://huggingface.co/w4ashabii)
- **License:** MIT

### Dataset Sources

- **Source text:** Nepali dictionary, 10th edition (https://archive.org/details/nepali-brihat-sabdakosh-nepali-dictionary/page/n11/mode/2up)

## Dataset Structure

53,119 rows, single `train` split, CSV format with 5 columns:

| Column | Description |
|---|---|
| `Word` | Raw surface form (e.g. `अँकाइनु`) |
| `Root` | The morphological root (e.g. `अँका`) |
| `Affix` | The affix, or `null` if the word has no segmentable affix |
| `Segmented` | `Root`+`Affix` joined with `+`, or equal to `Word` for unsegmented entries |
| `Source` | How the row was produced — one of `dictionary`, `heuristic`, `none` |

`Source` values:
- **`dictionary`** — root/affix pair taken directly from a dictionary entry
- **`heuristic`** — root/affix inferred by the rule engine from a related dictionary entry, not itself a dictionary headword
- **`none`** — the word has no affix; `Root` == `Word` == `Segmented`

## Dataset Creation

### Curation Rationale

Built to support morpheme-boundary prediction for Nepali, particularly
generalization to **out-of-vocabulary (OOV) roots** not seen during
training — see the companion fine-tuned model,
[`w4ashabii/Byt5_OOV_finetuned`](https://huggingface.co/w4ashabii/Byt5_OOV_finetuned).

### Source Data

#### Data Collection and Processing

1. The 10th edition of the Nepali dictionary was scraped for headwords and
   their listed inflected/derived forms.(https://archive.org/details/nepali-brihat-sabdakosh-nepali-dictionary/page/n11/mode/2up)
2. The scraped entries were parsed into `(Word, Root)` pairs.
3. A heuristic rule engine extended coverage beyond literal dictionary
   entries — inferring root/affix splits for related surface forms sharing
   a dictionary-attested root (e.g. other inflections of a verb root that
   itself appears in the dictionary).
4. Each resulting row was tagged with its `Source` (`dictionary`,
   `heuristic`, or `none` for unsegmented words) to keep the two processes
   distinguishable downstream.

#### Who are the source data producers?

Nepali dictionary (10th edition) editors/lexicographers, via automated
scraping and heuristic parsing by the dataset curator.


### Recommendations

If training a model for OOV generalization on this data, split by **Root**
(not by `Word`) to avoid leaking the same root across train/val/test via
different inflections of it.

## Citation

If you use this dataset, please cite the Nepali dictionary, Sabdakosh 2075 10th edition,
as the underlying source, along with this repository.

**BibTeX:**
```bibtex
@misc{root_affix_dictionary,
  author = {w4ashabii},
  title = {Nepali Root-Affix Dictionary},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/w4ashabii/root_affix_dictionary}}
}
```