Simplifying Persian NLP for Modern Applications

Shekar (meaning 'sugar' in Persian) is an open-source Python library for Persian natural language processing, named after the influential satirical story "فارسی شکر است" (Persian is Sugar) published in 1921 by Mohammad Ali Jamalzadeh. The story became a cornerstone of Iran's literary renaissance, advocating for accessible yet eloquent expression. Shekar embodies this philosophy in its design and development.

It provides tools for text preprocessing, tokenization, part-of-speech(POS) tagging, named entity recognition(NER), embeddings, spell checking, and more. With its modular pipeline design, Shekar makes it easy to build reproducible workflows for both research and production applications.

📖 Documentation: https://lib.shekar.io/

Table of Contents

• Installation
  • CPU Installation (All Platforms)
  • GPU Acceleration (NVIDIA CUDA)
• Preprocessing
  • Normalizer
  • Customization
• Tokenization
  • WordTokenizer
  • SentenceTokenizer
• Embeddings
  • Word Embeddings
  • Contextual Embeddings
• Stemming
• Lemmatization
• Part-of-Speech Tagging
• Named Entity Recognition (NER)
• Sentiment Analysis
• Toxicity Detection
• Keyword Extraction
• Spell Checking
• WordCloud
• Command-Line Interface (CLI)
• Download Models

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)

bash $ 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

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

Preprocessing

Normalizer

The built-in Normalizer class provides a ready-to-use pipeline that combines the most common filters and normalization steps, offering a default configuration that covers the majority of use cases.

```python 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)) ```

shell «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.

Customization

For advanced customization, Shekar offers a modular and composable framework for text preprocessing. It includes components such as filters, normalizers, and maskers, which can be applied individually or flexibly combined using the Pipeline class with the | operator.

You can combine any of the preprocessing components using the | operator:

```python from shekar.preprocessing import EmojiRemover, PunctuationRemover

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

shell 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.

```python from shekar import WordTokenizer

tokenizer = WordTokenizer()

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

shell ["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:

```python 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) ```

output 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.

```python from shekar.embeddings import WordEmbedder

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

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

similarwords = embedder.mostsimilar("کتاب", topn=5) print(similarwords) ```

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.

```python 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.

```python from shekar import Stemmer

stemmer = Stemmer()

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

output نوه کتاب 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.

```python from shekar import Lemmatizer

lemmatizer = Lemmatizer()

print(lemmatizer("رفتند")) print(lemmatizer("کتاب‌ها")) print(lemmatizer("خانه‌هایی")) print(lemmatizer("گفته بوده‌ایم")) ```

output رفت/رو کتاب خانه گفت/گو

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:

```python 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}") ```

output 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:

```python 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() normalizedtext = normalizer(inputtext)

albertner = NER() entities = albertner(normalized_text)

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

output 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:

```python from shekar import SentimentClassifier

sentiment_classifier = SentimentClassifier()

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

output ('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.

```python from shekar.toxicity import OffensiveLanguageClassifier

offensive_classifier = OffensiveLanguageClassifier()

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

output ('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.

```python from shekar import KeywordExtractor

extractor = KeywordExtractor(maxlength=2, topn=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) output 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.

```python from shekar import SpellChecker

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

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

WordCloud

The WordCloud class offers an easy way to create visually rich Persian word clouds. It supports reshaping and right-to-left rendering, Persian fonts, color maps, and custom shape masks for accurate and elegant visualization of word frequencies.

```python import requests from collections import Counter

from shekar 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) htmlcontent = response.text cleantext = preprocessingpipeline(htmlcontent)

wordtokenizer = WordTokenizer() tokens = wordtokenizer(clean_text)

word_freqs = Counter(tokens)

wordCloud = WordCloud( mask="Iran", width=640, height=480, maxfontsize=220, minfontsize=6, bgcolor="white", contourcolor="black", contourwidth=5, colormap="greens", )

if shows disconnect words, try again with bidi_reshape=True

image = wordCloud.generate(wordfreqs, bidireshape=False) image.show() ```

Command-Line Interface (CLI)

Shekar includes a command-line interface (CLI) for quick text processing and visualization.
You can normalize Persian text or generate wordclouds directly from files or inline strings.

Usage

console shekar [COMMAND] [OPTIONS]

Examples

```console

Normalize a text file and save output

shekar normalize -i ./corpus.txt -o ./normalized_corpus.txt

Normalize inline text

shekar normalize -t "درود پرودگار بر ایران و ایرانی" ```

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)|

With ❤️ for IRAN