Udmurt POS-Tagger

Python library for part-of-speech (POS) tagging of Udmurt language sentences using a bidirectional LSTM neural network model.

Features

• Trained on the udmurtNLP/zerpal-pos-tagging dataset
• Supports 17 POS tags
• Achieves 91.83% accuracy on the test set
• Handles out-of-vocabulary words using special embeddings

Installation

git clone https://github.com/codemurt/udmurt_postagger.git
cd udmurt_postagger
pip install .

Usage

from udmurt_postagger import UdmurtPOSTagger

tagger = UdmurtPOSTagger()
sentence = "Мон мынӥсько школае .".split()
tags = tagger.predict(sentence)

print("Token -> POS Tag")
for token, tag in zip(sentence, tags):
    print(f"{token} -> {tag}")

Model Architecture

The model uses a bidirectional LSTM architecture with the following layers:

• Embedding Layer: 128-dimensional word embeddings
• Bidirectional LSTM: 256 units (512 total output dimensions)
• TimeDistributed Dense Layer: 18 units (17 tags + padding)
• Activation: Softmax activation for tag probability distribution

Total Parameters: 3,060,754 (11.68 MB)

Performance

Trained on 9,913 sentences and evaluated on 2,479 sentences:

Tag	Precision	Recall	F1-Score	Support
N	0.8740	0.9058	0.8896	5957
V	0.8994	0.9062	0.9028	6261
PRO	0.9889	0.9797	0.9843	1724
CNJ	0.7901	0.8081	0.7990	787
PART	0.9162	0.8961	0.9061	1916
NUM	0.9787	0.9200	0.9485	250
IMIT	0.7778	0.4038	0.5316	104
ADV	0.8628	0.8332	0.8477	1283
O	0.9998	0.9956	0.9977	6199
INTRJ	0.9500	0.8636	0.9048	66
ADJPRO	0.9769	0.9270	0.9513	137
ADJ	0.8024	0.7357	0.7676	927
PREDIC	0.9200	0.9718	0.9452	71
POST	0.9343	0.9309	0.9326	275
ADVPRO	0.9972	0.9757	0.9863	370
PREP	0.5000	0.7500	0.6000	4
PARENTH	0.8889	0.9275	0.9078	69

Overall Accuracy: 91.83%
Weighted F1-Score: 91.79%

Training Details

• Optimizer: Adam (learning rate 0.001)
• Batch Size: 128
• Epochs: 40
• Padding: Post-padding to maximum sequence length (89 tokens)
• Special Tokens:
  • -PAD- for padding
  • -OOV- for out-of-vocabulary words

File Descriptions

• udmurt_pos_tagger_model.h5: Trained Keras model
• udmurt_pos_word2index.pkl: Word-to-index mapping dictionary
• udmurt_pos_index2tag.pkl: Index-to-tag mapping dictionary