Simplifying Persian NLP for Modern Applications

Shekar is an open-source Python library for Persian natural language processing, inspired by the satirical story فارسی شکر است (Persian is Sugar) by Mohammad Ali Jamalzadeh. Reflecting its emphasis on clear and accessible language, Shekar provides fast, modular tools for Persian text processing, including normalization, tokenization, POS tagging, NER, embeddings, and spell checking, enabling reproducible workflows for both research and production.

Why Shekar?

• Advanced text normalization: Built for the complexity of Persian text.
• Blazing fast and production-ready: Optimized for large-scale processing and real-time use.
• Modular and highly customizable: Independent, composable components for flexible NLP pipelines.
• Lightweight and efficient: Minimal dependencies and small models for fast CPU inference.
• Reliable and well-tested: Backed by hundreds of test cases with 95%+ code coverage.

Installation

You can install Shekar with pip. By default, the CPU runtime of ONNX is included, which works on all platforms.

CPU Installation (All Platforms)

$ pip install shekar

This works on Windows, Linux, and macOS (including Apple Silicon M1/M2/M3).

GPU Acceleration (NVIDIA CUDA)

If you have an NVIDIA GPU and want hardware acceleration, you need to replace the CPU runtime with the GPU version.

Prerequisites

• NVIDIA GPU with CUDA support
• Appropriate CUDA Toolkit installed
• Compatible NVIDIA drivers

$ pip install shekar && pip uninstall -y onnxruntime && pip install onnxruntime-gpu

Preprocessing

Normalizer

The built-in Normalizer class provides a ready-to-use, opinionated normalization pipeline for Persian text. It combines the most common and error-prone normalization steps into a single component, covering the majority of real-world use cases such as web text, social media, OCR output, and mixed informal–formal writing.

Most importantly, the normalization rules in Shekar strictly follow the official guidelines of Academy of Persian Language and Literature (فرهنگستان زبان و ادب فارسی). This makes the output suitable not only for NLP pipelines, but also for linguistically correct and publishable Persian text.

from shekar import Normalizer

normalizer = Normalizer()

text = "«فارسی شِکَر است» نام داستان ڪوتاه طنز    آمێزی از محمد علی جمالــــــــزاده ی گرامی می   باشد که در سال 1921 منتشر  شده است و آغاز   ڱر تحول بزرگی در ادَبێات معاصر ایران بۃ شمار میرود."
print(normalizer(text))

# نرمال‌سازی نویسه‌های گفتاری و روزمره
text = normalizer("می دونی که نمیخاستم ناراحتت کنم.اما خونه هاشون خیلی گرون تر شده")
print(text)

# نرمال‌سازی واژه‌های مرکب و افعال پیشوندی! 
text = normalizer("یک کار آفرین نمونه و سخت کوش ، پیروز مندانه از پس دشواری ها برخواهدآمد.")
print(text) 

«فارسی شکر است» نام داستان کوتاه طنزآمیزی از محمد‌علی جمالزاده‌ی گرامی می‌باشد که در سال ۱۹۲۱ منتشر شده‌است و آغازگر تحول بزرگی در ادبیات معاصر ایران به شمار می‌رود.

می‌دونی که نمی‌خاستم ناراحتت کنم. اما خونه‌هاشون خیلی گرون‌تر شده

یک کارآفرین نمونه و سخت‌کوش، پیروزمندانه از پس دشواری‌ها بر خواهد آمد.

Customization

Shekar is built around a modular and composable preprocessing framework that allows fine-grained control over each step of text processing. Preprocessing is implemented as small, independent operators such as filters, normalizers, and maskers, which can be used on their own or combined into flexible pipelines.

Pipelines are constructed using the Pipeline abstraction and composed with the | operator, making preprocessing logic explicit, readable, and easy to customize. Any operator from the full list of preprocessing components can be freely combined.

For example, the following pipeline is functionally equivalent to the default normalizer:

from shekar.preprocessing import (
    PunctuationNormalizer,
    AlphabetNormalizer,
    DigitNormalizer,
    SpacingNormalizer,
    RemoveDiacritics,
    RepeatedLetterNormalizer,
    ArabicUnicodeNormalizer,
    YaNormalizer,
)

normalizer = (
            AlphabetNormalizer()
            | ArabicUnicodeNormalizer()
            | DigitNormalizer()
            | PunctuationNormalizer()
            | RemoveDiacritics()
            | RepeatedLetterNormalizer()
            | SpacingNormalizer()
            | YaNormalizer(style="joda")
        )

Operators can also be composed for lightweight, task-specific preprocessing. For example, removing emojis and punctuation:

from shekar.preprocessing import EmojiRemover, PunctuationRemover

text = "ز ایران دلش یاد کرد و بسوخت! 🌍🇮🇷"
pipeline = EmojiRemover() | PunctuationRemover()
output = pipeline(text)
print(output)

ز ایران دلش یاد کرد و بسوخت

Tokenization

WordTokenizer

The WordTokenizer class in Shekar is a simple, rule-based tokenizer for Persian that splits text based on punctuation and whitespace using Unicode-aware regular expressions.

from shekar import WordTokenizer

tokenizer = WordTokenizer()

text = "چه سیب‌های قشنگی! حیات نشئهٔ تنهایی است."
tokens = list(tokenizer(text))
print(tokens)

["چه", "سیب‌های", "قشنگی", "!", "حیات", "نشئهٔ", "تنهایی", "است", "."]

SentenceTokenizer

The SentenceTokenizer class is designed to split a given text into individual sentences. This class is particularly useful in natural language processing tasks where understanding the structure and meaning of sentences is important. The SentenceTokenizer class can handle various punctuation marks and language-specific rules to accurately identify sentence boundaries.

Below is an example of how to use the SentenceTokenizer:

from shekar.tokenization import SentenceTokenizer

text = "هدف ما کمک به یکدیگر است! ما می‌توانیم با هم کار کنیم."
tokenizer = SentenceTokenizer()
sentences = tokenizer(text)

for sentence in sentences:
    print(sentence)

هدف ما کمک به یکدیگر است!
ما می‌توانیم با هم کار کنیم.

Embeddings

Shekar offers two main embedding classes:

• WordEmbedder: Provides static word embeddings using pre-trained FastText models.
• ContextualEmbedder: Provides contextual embeddings using a fine-tuned ALBERT model.

Both classes share a consistent interface:

• embed(text) returns a NumPy vector.
• transform(text) is an alias for embed(text) to integrate with pipelines.

Word Embeddings

WordEmbedder supports two static FastText models:

• fasttext-d100: A 100-dimensional CBOW model trained on Persian Wikipedia.
• fasttext-d300: A 300-dimensional CBOW model trained on the large-scale Naab dataset.

from shekar.embeddings import WordEmbedder

embedder = WordEmbedder(model="fasttext-d100")

embedding = embedder("کتاب")
print(embedding.shape)

similar_words = embedder.most_similar("کتاب", top_n=5)
print(similar_words)

Contextual Embeddings

ContextualEmbedder uses an ALBERT model trained with Masked Language Modeling (MLM) on the Naab dataset to generate high-quality contextual embeddings. The resulting embeddings are 768-dimensional vectors representing the semantic meaning of entire phrases or sentences.

from shekar.embeddings import ContextualEmbedder

embedder = ContextualEmbedder(model="albert")

sentence = "کتاب‌ها دریچه‌ای به جهان دانش هستند."
embedding = embedder(sentence)
print(embedding.shape)  # (768,)

Stemming

The Stemmer is a lightweight, rule-based reducer for Persian word forms. It trims common suffixes while respecting Persian orthography and Zero Width Non-Joiner usage. The goal is to produce stable stems for search, indexing, and simple text analysis without requiring a full morphological analyzer.

from shekar import Stemmer

stemmer = Stemmer()

print(stemmer("نوه‌ام"))
print(stemmer("کتاب‌ها"))
print(stemmer("خانه‌هایی"))
print(stemmer("خونه‌هامون"))

نوه
کتاب
خانه
خانه

Lemmatization

The Lemmatizer maps Persian words to their base dictionary form. Unlike stemming, which only trims affixes, lemmatization uses explicit verb conjugation rules, vocabulary lookups, and a stemmer fallback to ensure valid lemmas. This makes it more accurate for tasks like part-of-speech tagging, text normalization, and linguistic analysis where the canonical form of a word is required.

from shekar import Lemmatizer

lemmatizer = Lemmatizer()

# ریشه‌یابی افعال
print(lemmatizer("رفتند"))
print(lemmatizer("گفته بوده‌ایم"))

# ریشه‌یابی واژه‌ها
print(lemmatizer("کتاب‌ها"))
print(lemmatizer("خانه‌هایی"))
print(lemmatizer("خونه‌هامون"))

# ریشه‌یابی افعال پیشوندی
print(lemmatizer("بر نخواهم گشت"))
print(lemmatizer("برنمی‌دارم"))

رفت/رو
گفت/گو
کتاب
خانه
خانه
برگشت/برگرد
برداشت/بردار

Part-of-Speech Tagging

The POSTagger class provides part-of-speech tagging for Persian text using a transformer-based model (default: ALBERT). It returns one tag per word based on Universal POS tags (following the Universal Dependencies standard).

Example usage:

from shekar import POSTagger

pos_tagger = POSTagger()
text = "نوروز، جشن سال نو ایرانی، بیش از سه هزار سال قدمت دارد و در کشورهای مختلف جشن گرفته می‌شود."

result = pos_tagger(text)
for word, tag in result:
    print(f"{word}: {tag}")

نوروز: PROPN
،: PUNCT
جشن: NOUN
سال: NOUN
نو: ADJ
ایرانی: ADJ
،: PUNCT
بیش: ADJ
از: ADP
سه: NUM
هزار: NUM
سال: NOUN
قدمت: NOUN
دارد: VERB
و: CCONJ
در: ADP
کشورهای: NOUN
مختلف: ADJ
جشن: NOUN
گرفته: VERB
می‌شود: VERB
.: PUNCT

Named Entity Recognition (NER)

The NER module offers a fast, quantized Named Entity Recognition pipeline using a fine-tuned ALBERT model. It detects common Persian entities such as persons, locations, organizations, and dates. This model is designed for efficient inference and can be easily combined with other preprocessing steps.

Example usage:

from shekar import NER
from shekar import Normalizer

input_text = (
    "شاهرخ مسکوب به سالِ ۱۳۰۴ در بابل زاده شد و دوره ابتدایی را در تهران و در مدرسه علمیه پشت "
    "مسجد سپهسالار گذراند. از کلاس پنجم ابتدایی مطالعه رمان و آثار ادبی را شروع کرد. از همان زمان "
    "در دبیرستان ادب اصفهان ادامه تحصیل داد. پس از پایان تحصیلات دبیرستان در سال ۱۳۲۴ از اصفهان به تهران رفت و "
    "در رشته حقوق دانشگاه تهران مشغول به تحصیل شد."
)

normalizer = Normalizer()
normalized_text = normalizer(input_text)

albert_ner = NER()
entities = albert_ner(normalized_text)

for text, label in entities:
    print(f"{text} → {label}")

شاهرخ مسکوب → PER
سال ۱۳۰۴ → DAT
بابل → LOC
دوره ابتدایی → DAT
تهران → LOC
مدرسه علمیه → LOC
مسجد سپهسالار → LOC
دبیرستان ادب اصفهان → LOC
در سال ۱۳۲۴ → DAT
اصفهان → LOC
تهران → LOC
دانشگاه تهران → ORG
فرانسه → LOC

Sentiment Analysis

The SentimentClassifier module enables automatic sentiment analysis of Persian text using transformer-based models. It currently supports the AlbertBinarySentimentClassifier, a lightweight ALBERT model fine-tuned on Snapfood dataset to classify text as positive or negative, returning both the predicted label and its confidence score.

Example usage:

from shekar import SentimentClassifier

sentiment_classifier = SentimentClassifier()

print(sentiment_classifier("سریال قصه‌های مجید عالی بود!"))
print(sentiment_classifier("فیلم ۳۰۰ افتضاح بود!"))

('positive', 0.9923112988471985)
('negative', 0.9330866932868958)

Toxicity Detection

The toxicity module currently includes a Logistic Regression classifier trained on TF-IDF features extracted from the Naseza (ناسزا) dataset, a large-scale collection of Persian text labeled for offensive and neutral language. The OffensiveLanguageClassifier processes input text to determine whether it is neutral or offensive, returning both the predicted label and its confidence score.

from shekar.toxicity import OffensiveLanguageClassifier

offensive_classifier = OffensiveLanguageClassifier()

print(offensive_classifier("زبان فارسی میهن من است!"))
print(offensive_classifier("تو خیلی احمق و بی‌شرفی!"))

('neutral', 0.7651197910308838)
('offensive', 0.7607775330543518)

Keyword Extraction

The shekar.keyword_extraction module provides tools for automatically identifying and extracting key terms and phrases from Persian text. These algorithms help identify the most important concepts and topics within documents.

from shekar import KeywordExtractor

extractor = KeywordExtractor(max_length=2, top_n=10)

input_text = (
    "زبان فارسی یکی از زبان‌های مهم منطقه و جهان است که تاریخچه‌ای کهن دارد. "
    "زبان فارسی با داشتن ادبیاتی غنی و شاعرانی برجسته، نقشی بی‌بدیل در گسترش فرهنگ ایرانی ایفا کرده است. "
    "از دوران فردوسی و شاهنامه تا دوران معاصر، زبان فارسی همواره ابزار بیان اندیشه، احساس و هنر بوده است. "
)

keywords = extractor(input_text)

for kw in keywords:
    print(kw)

فرهنگ ایرانی
گسترش فرهنگ
ایرانی ایفا
زبان فارسی
تاریخچه‌ای کهن

Spell Checking

The SpellChecker class provides simple and effective spelling correction for Persian text. It can automatically detect and fix common errors such as extra characters, spacing mistakes, or misspelled words. You can use it directly as a callable on a sentence to clean up the text, or call suggest() to get a ranked list of correction candidates for a single word.

from shekar import SpellChecker

spell_checker = SpellChecker()
print(spell_checker("سسلام بر ششما ددوست من"))
print(spell_checker.suggest("درود"))

سلام بر شما دوست من
['درود', 'درصد', 'ورود', 'درد', 'درون']

WordCloud

The WordCloud class provides a convenient interface for generating Persian word clouds with correct shaping, directionality, and typography. It is specifically designed to work with right-to-left Persian text and integrates seamlessly with Shekar’s normalization utilities to produce visually accurate and linguistically correct results.

The WordCloud functionality depends on visualization libraries that are not installed by default. To enable this feature, install Shekar with the optional visualization dependencies:

$ pip install 'shekar[viz]'

Example usage:

import requests
from collections import Counter

from shekar.visualization import WordCloud
from shekar import WordTokenizer
from shekar.preprocessing import (
  HTMLTagRemover,
  PunctuationRemover,
  StopWordRemover,
  NonPersianRemover,
)
preprocessing_pipeline = HTMLTagRemover() | PunctuationRemover() | StopWordRemover() | NonPersianRemover()


url = f"https://shahnameh.me/p.php?id=F82F6CED"
response = requests.get(url)
html_content = response.text
clean_text = preprocessing_pipeline(html_content)

word_tokenizer = WordTokenizer()
tokens = word_tokenizer(clean_text)

word_freqs = Counter(tokens)

wordCloud = WordCloud(
        mask="Iran",
        width=640,
        height=480,
        max_font_size=220,
        min_font_size=6,
        bg_color="white",
        contour_color="black",
        contour_width=5,
        color_map="greens",
    )

# if shows disconnect words, try again with bidi_reshape=True
image = wordCloud.generate(word_freqs, bidi_reshape=False)
image.show()

Download Models

If Shekar Hub is unavailable, you can manually download the models and place them in the cache directory at home/[username]/.shekar/

Model Name	Download Link
FastText Embedding d100	Download (50MB)
FastText Embedding d300	Download (500MB)
SentenceEmbedding	Download (60MB)
POS Tagger	Download (38MB)
NER	Download (38MB)
Sentiment Classifier	Download (38MB)
Offensive Language Classifier	Download (8MB)
AlbertTokenizer	Download (2MB)

---

Citation

If you find Shekar useful in your research, please consider citing the following paper:

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

With ❤️ for IRAN