khmer-dictionary / README.md
phonsobon's picture
Add g2p config (word->pronunciation, train/val/test); fix pronunciation typos
2afb0c6 verified
|
Raw History Blame Contribute Delete
9.54 kB
---
language:
- km
license: other
license_name: nckl-khmer-dictionary-2022
pretty_name: Khmer Dictionary 2022 (NCKL)
size_categories:
- 10K<n<100K
task_categories:
- text-generation
- fill-mask
tags:
- khmer
- dictionary
- lexicon
- monolingual-dictionary
configs:
- config_name: entries
default: true
data_files:
- split: train
path: data/entries.parquet
- config_name: senses
data_files:
- split: train
path: data/senses.parquet
- config_name: g2p
data_files:
- split: train
path: data/g2p/train.parquet
- split: validation
path: data/g2p/validation.parquet
- split: test
path: data/g2p/test.parquet
---
# Khmer Dictionary 2022 (វចនានុក្រមខ្មែរ ២០២២)
A clean, machine-readable version of the **Khmer Dictionary 2022** from the
**National Council of Khmer Language (NCKL), Royal Academy of Cambodia**.
It is a monolingual dictionary: the headwords, definitions and examples are all in Khmer.
| | count |
|---|---:|
| Entries (words with their own definition) | 41,087 |
| Senses (individual meanings) | 44,698 |
| Unique headwords | 25,285 |
| Sub-entries (compounds/derived words under a headword) | 13,477 |
| Unique words (headwords + sub-entries) | 38,674 |
## Quick start
```python
from datasets import load_dataset
# One row per word, with all its meanings nested in `senses`
entries = load_dataset("phonsobon/khmer-dictionary", "entries", split="train")
# One flat row per meaning, which is easier for pandas, search and training
senses = load_dataset("phonsobon/khmer-dictionary", "senses", split="train")
df = senses.to_pandas()
df[df.word == "កក"][["word", "pronunciation", "part_of_speech_en", "definition", "examples"]]
```
## Configurations
### `entries` (default): one row per word
```json
{
"entry_id": 3,
"word": "ក",
"headword": "ក",
"homograph": 3,
"is_subentry": false,
"pronunciation": "ក",
"pronunciation_note": "",
"etymology": "",
"senses": [
{"sense_no": 1, "sense_label": "១)", "part_of_speech": "ន.", "part_of_speech_en": "noun",
"definition": "អវយវៈដែលតពីក្បាលទៅស្មាឬទៅខ្លួននៃមនុស្សសត្វ",
"examples": "កមនុស្ស, កមាន់។", "notes": ""},
{"sense_no": 2, "sense_label": "២)", "part_of_speech": "ន.", "part_of_speech_en": "noun",
"definition": "កន្លែងដែលតជាប់ពីមាត់ដបជាដើមចុះទៅទល់នឹងក្អេងក៏ហៅ ក ដែរ",
"examples": "កដប, កក្អម។", "notes": ""}
]
}
```
### `senses`: one row per meaning (flat)
The `senses` config has the same fields, flattened. Every row repeats its word's
`word`, `pronunciation`, `etymology`, etc., so each row makes sense on its own.
It also has `sense_id`, the original row id. A CSV copy is in `data/senses.csv`,
and a plain list of every word is in `data/words.txt`.
### `g2p`: word → pronunciation pairs for grapheme-to-phoneme training
Clean spelling → pronunciation pairs, ready for training a Khmer G2P,
transliteration or spelling-check model. It is split by word, so all readings
of a word are in the same split (90 / 5 / 5).
| split | pairs |
|---|---:|
| train | 34,807 |
| validation | 1,942 |
| test | 1,934 |
```json
{"word": "កករចិត្ត", "pronunciation": "ក-ក-ចិត", "entry_ids": [20], "syllables": 3}
```
| Field | Description |
|---|---|
| `word` | Spelling (Khmer script only) |
| `pronunciation` | Reading in Khmer script, syllables separated by `-`. Remove the `-` if you only need the phoneme string |
| `entry_ids` | The dictionary entries this pair comes from (homographs share one row) |
| `syllables` | Number of syllables |
What is left out compared with the full dictionary:
- **163 proverbs and idioms with no pronunciation** (e.g. `កង្កែបអាចម៍គោហ៊ានលោតខាំពស់វែក`). They are still in `entries` / `senses`.
- **Bound/combining forms** such as `កម្ពុជ—`. They are not standalone words, and their reading changes inside compounds.
- **Phrases with spaces, `...` or abbreviations** in brackets.
Alternative readings (`a ឬ b`) become separate rows, and duplicate
(word, pronunciation) pairs are merged.
```python
g2p = load_dataset("phonsobon/khmer-dictionary", "g2p")
g2p["train"][0] # {'word': 'ក', 'pronunciation': 'ក', ...}
```
## Fields
| Field | Type | Description |
|---|---|---|
| `entry_id` | int | ID of the entry (stable, from the source database) |
| `sense_id` | int | *(senses only)* ID of this sense row in the source |
| `word` | string | **The word being defined.** For a sub-entry this is the compound (e.g. `កកកុញ`); otherwise it is the headword |
| `headword` | string | The main dictionary headword this entry belongs to (e.g. `កក`) |
| `homograph` | int / null | Homograph number when several unrelated headwords share a spelling (ក¹, ក², ក³ …) |
| `is_subentry` | bool | `true` if `word` is a compound/derived word listed under `headword` |
| `pronunciation` | string | Syllable-by-syllable reading in Khmer script, syllables joined by `-`. Alternative readings are joined with ` ឬ ` ("or") |
| `pronunciation_note` | string | Reading note, e.g. `ជើង ត អានជា ដ` (subscript ត is read as ដ) |
| `etymology` | string | Origin of the word, e.g. `(បា.)` = from Pali (see abbreviations below) |
| `sense_no` | int | Order of this meaning within the entry (1, 2, 3 …) |
| `sense_label` | string | Original sense marker in the dictionary: `១)`, `២)` … for numbered meanings, `a`, `b` … for sub-meanings, empty if there is only one |
| `part_of_speech` | string | Part of speech as printed in Khmer (e.g. `ន.`, `គុ.កិ., គុ.`) |
| `part_of_speech_en` | string | The same in English (e.g. `noun`, `adverb, adjective`) |
| `definition` | string | The definition |
| `examples` | string | Usage examples (comma-separated phrases or sentences, ending with `។`) |
| `notes` | string | Extra notes such as cross-references, e.g. `(ម.ព. ត្រង¹ ផង)។` = "see also ត្រង¹" |
## Abbreviations
**Parts of speech** (`part_of_speech` → `part_of_speech_en`)
| Khmer | Full form | English |
|---|---|---|
| ន. | នាម | noun |
| កិ. | កិរិយាសព្ទ | verb |
| គុ. | គុណនាម | adjective |
| គុ.កិ. | គុណកិរិយា | adverb |
| សព្វ. | សព្វនាម | pronoun |
| ប. | បុព្វបទ | preposition |
| ឈ. | ឈ្នាប់ | conjunction |
| ឧ. | ឧទានសព្ទ | interjection |
| ធ. | ធ្នាក់ | particle |
**Etymology** (`etymology`)
| Abbr. | Language |
|---|---|
| បា. | Pali |
| សំ. | Sanskrit |
| សំ.បា. / បា.សំ. | Sanskrit and Pali |
| បារ. | French (បារាំង) |
| ស. | Thai/Siamese (សៀម) |
| ចិ. | Chinese |
| អ. | English |
| យ. | Vietnamese (យួន) |
| ល. | Lao |
Inside etymologies, `<` and `>` show derivation (`A > B` = A became B), and
`អ.ថ.` means "read as".
**Common markers in definitions and notes**
| Marker | Meaning |
|---|---|
| `ម.ព.` | see (cross-reference) |
| `(ព.ប្រ.)` | figurative usage |
| `(ព.កា.)` | poetic/literary usage |
| `(គណិត.)`, `(គីមី.)` … | domain label: mathematics, chemistry … |
| `ក¹`, `ក²` | homograph number of the referenced word |
## How this was cleaned
The source export was a flat table where the 2nd, 3rd … meanings of a word had
blank word columns and some columns had misleading names. The changes:
- **Filled every row in.** Each meaning now carries its word, pronunciation, etymology and part of speech.
- **Renamed the columns.** The source `sense` column (which actually held the compound word) became `word`, `poly` became `sense_label`, and `ref` became `entry_id`.
- **Converted the homograph markup.** The internal markup `{¥¥}១{££}` is now a superscript number (`¹`).
- **Cleaned the pronunciations.** The `[ ]` brackets are removed, and reading notes moved to `pronunciation_note`.
- **Trimmed the definitions.** The trailing ` :`, which only pointed at the examples, is removed.
- **Added English POS labels.** These are in `part_of_speech_en`.
- **Dropped one column.** The undocumented `remark` flag was removed.
- **Fixed pronunciation typos.** Stray `´´`, `:` typed in place of `ៈ`, and doubled `[[` (6 entries).
- **Left proverbs and idioms empty.** Unlike some other copies of this dictionary, the 163 proverbs/idioms with no pronunciation in the source stay empty instead of borrowing the pronunciation of the row above.
The Khmer text is otherwise unchanged. Zero-width joiners and non-joiners
(e.g. in `អ៊‌ីចឹង`) are kept on purpose: Khmer uses them to control how
characters combine. A few etymologies of Thai loanwords contain
private-use glyphs from the original font.
The build script is `scripts/build_hf_dataset.py`.
## Source and license
Content © National Council of Khmer Language (ក្រុមប្រឹក្សាជាតិភាសាខ្មែរ),
Royal Academy of Cambodia, *Khmer Dictionary 2022*. This dataset is a
reformatted copy for research and language-technology use. All rights in the
dictionary content remain with NCKL. Check NCKL's terms before any commercial use
or redistribution.