khmer-nlp-kcc 0.2.7

Khmer word segmentation, POS tagging, NER, and more — powered by a custom Khmer language model

khmer-nlp-kcc

Khmer NLP toolkit — word segmentation, POS tagging, sentiment polarity, and grapheme-to-phoneme conversion.

Powered by a custom Khmer language model. The NLP checkpoint is downloaded automatically from HuggingFace Hub on first use.

Installation

pip install khmer-nlp-kcc

Quick start

from khmer_nlp import KhmerNLP

nlp = KhmerNLP()  # NLP checkpoint downloads on first call

# Word segmentation
nlp.segment("គាត់និយាយថា៖ការសិក្សានាំមកនូវចំណេះដឹងតើឯកភាពទេ?")
# → "គាត់ និយាយ ថា ៖ ការសិក្សា នាំ មក នូវ ចំណេះដឹង តើ ឯកភាព ទេ ?"

# Word segmentation with sub-word markers
nlp.segment("គាត់និយាយថា៖ការសិក្សានាំមកនូវចំណេះដឹងតើឯកភាពទេ?", markers=True)

# POS tagging
nlp.pos("គាត់និយាយថា៖ការសិក្សានាំមកនូវចំណេះដឹងតើឯកភាពទេ?")
# → [{"word": "គាត់", "label": "PRO"}, {"word": "និយាយ", "label": "VB"}, ...]

# Nova POS tagging
nlp.nova_pos("គាត់និយាយថា៖ការសិក្សានាំមកនូវចំណេះដឹងតើឯកភាពទេ?")
# → [{"word": "គាត់", "label": "o"}, {"word": "និយាយ", "label": "v"}, ...]

# Sentiment polarity
nlp.polarity("គាត់និយាយថា៖ការសិក្សានាំមកនូវចំណេះដឹងតើឯកភាពទេ?")
# → {"label": "neutral", "confidence": ..., "scores": {...}}

# All tasks at once
nlp.analyze("គាត់និយាយថា៖ការសិក្សានាំមកនូវចំណេះដឹងតើឯកភាពទេ?")

Grapheme-to-phoneme (G2P)

The G2P checkpoint (g2p_final_trans.pt) is downloaded automatically from HuggingFace Hub on first use, just like the main model.

nlp = KhmerNLP()

# Phoneme sequence for a single word
nlp.g2p("ខ្មែរ")
# → ['kh', 'ae', '.', 'm', 'ae']

# Phoneme CER between two Khmer words
nlp.phoneme_cer("ខ្មែរ", "ខ្មែរភូមិ")
# → {
#     "word1": "ខ្មែរ",    "phones1": ["kh", "ae", ".", "m", "ae"],
#     "word2": "ខ្មែរភូមិ", "phones2": ["kh", "ae", ".", "m", "ae", ".", "ph", "uu", "m", ".", "m", "ɨ"],
#     "cer": 0.7
#   }

CER is computed at phoneme-token level (edit distance / number of phonemes in word1).

To use a local checkpoint instead:

nlp = KhmerNLP(g2p_checkpoint_path="/path/to/g2p_final_trans.pt")

Custom checkpoint

nlp = KhmerNLP(checkpoint_path="/path/to/your/model.pt")

Device selection

import torch
nlp = KhmerNLP(device=torch.device("cuda:0"))

Available methods

Method	Returns	Description
segment(text, *, markers=False)	str	Space-joined segmented words; markers=True adds sub-word markers (_)
pos(text)	list[dict]	Word + POS label
nova_pos(text)	list[dict]	Word + Nova POS label
polarity(text)	dict	Sentiment polarity
analyze(text)	dict	All tasks combined
g2p(word)	list[str]	Phoneme token sequence for a Khmer word
phoneme_cer(word1, word2)	dict	Phoneme CER between two Khmer words

Raw token-level outputs: seg_tokens(), pos_tokens().

License

MIT