shekar-ai/albert-base-v2-persian-pos-tagger

ALBERT Persian POS Tagger

Part-of-speech tagger for Persian using the 17-tag Universal Dependencies scheme. Fine-tuned from shekar-ai/albert-base-v2-persian-zwnj-naab-mlm on the Persian Universal Dependency Treebank (PerDT).

Task	Token classification (17 UD tags)
Parameters	11.4 M
Base model	shekar-ai/albert-base-v2-persian-zwnj-naab-mlm
Training data	PerDT (UD_Persian-PerDT)
License	MIT

Tags: ADJ, ADP, ADV, AUX, CCONJ, DET, INTJ, NOUN, NUM, PART, PRON, PROPN, PUNCT, SCONJ, VERB, X, _.

Usage

Shekar

from shekar import POSTagger

tagger = POSTagger()

for word, tag in tagger("کتاب‌ها دریچه‌ای به جهان دانش هستند."):
    print(f"{word}\t{tag}")

کتاب‌ها	NOUN
دریچه‌ای	NOUN
به	ADP
جهان	NOUN
دانش	NOUN
هستند	VERB
.	PUNCT

Transformers

The model was trained on word-level input with labels on the first subtoken of each word and no special tokens. Feeding raw text to the generic token-classification pipeline therefore produces subword-level noise — tokenize word by word and read the prediction of each word's first subtoken:

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

repo = "shekar-ai/albert-base-v2-persian-pos-tagger"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForTokenClassification.from_pretrained(repo).eval()

def tag(words):
    input_ids, word_ids = [], []
    for index, word in enumerate(words):
        subtokens = tokenizer.encode(word, add_special_tokens=False) or [tokenizer.unk_token_id]
        input_ids.extend(subtokens)
        word_ids.extend([index] * len(subtokens))

    with torch.no_grad():
        logits = model(input_ids=torch.tensor([input_ids])).logits[0]
    predictions = logits.argmax(-1).tolist()

    tags, previous = [], None
    for word_index, prediction in zip(word_ids, predictions):
        if word_index != previous:
            tags.append((words[word_index], model.config.id2label[prediction]))
            previous = word_index
    return tags

from shekar import WordTokenizer

for word, pos in tag(list(WordTokenizer()("شاهنامه اثر فردوسی در توس سروده شد."))):
    print(f"{word}\t{pos}")

شاهنامه	PROPN
اثر	NOUN
فردوسی	PROPN
در	ADP
توس	PROPN
سروده	VERB
شد	VERB
.	PUNCT

Training

Fine-tuned on the PerDT treebank: 26,196 training, 1,456 development, and 1,455 test sentences. Each word is tokenized independently; the gold tag is assigned to its first subtoken and remaining subtokens are masked out of the loss with -100. The checkpoint with the best macro-F1 was kept.

Hyperparameter	Value
Epochs	10
Batch size	8
Learning rate	2e-5
LR scheduler	Linear
Weight decay	0.01
Optimizer	AdamW
Max sequence length	512

Evaluation

Measured on the PerDT test split (1,455 sentences, 24,429 tokens):

Metric	Value
Accuracy	97.66
Macro F1	90.82

Per-tag scores:

Tag	Precision	Recall	F1	Support
ADJ	0.943	0.935	0.939	1,652
ADP	0.994	0.992	0.993	3,407
ADV	0.915	0.924	0.920	384
AUX	0.998	0.991	0.994	899
CCONJ	0.996	0.997	0.997	1,026
DET	0.972	0.982	0.977	491
INTJ	0.852	0.852	0.852	27
NOUN	0.974	0.971	0.972	8,219
NUM	0.970	0.980	0.975	293
PART	1.000	0.929	0.963	28
PRON	0.992	0.991	0.992	1,126
PROPN	0.873	0.907	0.890	1,111
PUNCT	0.999	0.999	0.999	2,141
SCONJ	0.990	0.979	0.985	632
VERB	0.996	0.998	0.997	2,696
X	0.000	0.000	0.000	1
_	1.000	0.993	0.997	296

Macro F1 is held down by the rare tags: INTJ (27 tokens) and X (1 token) contribute as much to the average as NOUN (8,219 tokens). The main substantive weakness is the NOUN/PROPN boundary, where Persian offers no capitalization cue.

Per-epoch validation during training

Epoch	Training loss	Validation loss	Accuracy	Macro F1
1	0.1121	0.1027	0.9699	0.8784
2	0.0726	0.0817	0.9753	0.9025
3	0.0563	0.0806	0.9763	0.9026
4	0.0396	0.0888	0.9761	0.9067
5	0.0280	0.1055	0.9765	0.9060
6	0.0124	0.1230	0.9766	0.9082
7	0.0078	0.1385	0.9763	0.9069
8	0.0042	0.1544	0.9758	0.9040
9	0.0017	0.1630	0.9763	0.9027
10	0.0012	0.1687	0.9763	0.9059

Citation

@article{Amirivojdan2025Shekar,
  author  = {Amirivojdan, Ahmad},
  title   = {{Shekar: A Python Toolkit for Persian Natural Language Processing}},
  journal = {Journal of Open Source Software},
  volume  = {10},
  number  = {114},
  pages   = {9128},
  year    = {2025},
  doi     = {10.21105/joss.09128},
  url     = {https://joss.theoj.org/papers/10.21105/joss.09128}
}