persian-sentiment

Persian Sentiment Analyzer

A Python library for sentiment analysis of Persian (Farsi) text, capable of classifying opinions as "recommended", "not_recommended", or "no_idea".

Features

• Text Preprocessing: Normalization, tokenization, stemming, and stopword removal for Persian text
• Word Embeddings: Built-in Word2Vec implementation for Persian language
• Sentiment Classification: Logistic Regression classifier trained on Persian sentiment data
• Model Persistence: Save and load trained models for future use
• Batch Processing: Analyze sentiment for multiple texts at once

Installation

pip install persian-sentiment-analyzer

Dependencies

• Python 3.6+

• hazm

• gensim

• scikit-learn

• numpy

• pandas

Usage

Basic Usage

from persian_sentiment_analyzer import SentimentAnalyzer

# Initialize with a pre-trained model
analyzer = SentimentAnalyzer(model_path="path/to/pretrained_model")

# Predict sentiment
result = analyzer.predict("این محصول بسیار عالی است")
print(result)  # Output: 'recommended'

Training Your Own Model

from persian_sentiment_analyzer import SentimentAnalyzer
import pandas as pd

# Load your dataset
data = pd.read_csv("persian_reviews.csv")
texts = data['text'].tolist()
labels = data['label'].values  # 0: not_recommended, 1: recommended, 2: no_idea

# Initialize analyzer
analyzer = SentimentAnalyzer()

# Preprocess and tokenize texts
tokenized_texts = [analyzer.preprocessor.preprocess_text(text) for text in texts]

# Train Word2Vec model
analyzer.train_word2vec(tokenized_texts, vector_size=100)

# Prepare feature vectors
X = np.array([analyzer.sentence_vector(tokens) for tokens in tokenized_texts])

# Train classifier
analyzer.train_classifier(X, labels)

# Save the trained model
analyzer.save_model("my_persian_model")

Batch Processing

from persian_sentiment_analyzer import predict_sentiments_for_file

# Process a CSV file containing Persian comments
results_summary = predict_sentiments_for_file(
    analyzer,
    input_file="comments.csv",
    output_file="results.csv",
    summary_file="summary.csv"
)

print(results_summary)

Model Architecture

1- Text Preprocessing:

• Normalization (Hazm)

• Tokenization

• Stemming

• Stopword removal

2- Feature Extraction:

• Word2Vec embeddings (100 dimensions)

• Sentence vectors (average of word vectors)

3- Classification:

• Logistic Regression with L2 regularization

Performance

The pre-trained model achieves the following performance on our test set:

Metric Value Accuracy 85.2% Precision 84.7% Recall 85.0% F1-score 84.8%

License This project is licensed under the MIT License - see the LICENSE file for details

Github : RezaGooner