PerSent 1.3.3

Persian Sentiment Analysis Toolkit

PerSent (Persian Sentiment Analyzer)

Introduction

PerSent is a practical Python library designed for Persian sentiment analysis. The name stands for "Persian Sentiment Analyzer". Currently in its early testing phase, the library provides basic functionality and is available on PyPI. Install it using:

pip install PerSent

Current capabilities include:

• Sentiment analysis of opinions/comments

• Emotion analysis of texts (happiness, sadness, anger, surprise, fear, disgust, calmness)

• Analysis of product/service reviews (recommended/not recommended/no idea)

• Both single-text and batch CSV processing

• Output displayed in terminal or saved to CSV with summary statistics

Initial repository that evolved into this library: Click here

We welcome user testing and feedback to improve the library. If you encounter bugs or have suggestions, please:

• Contribute

For installation issues due to dependency conflicts (especially with mingw-w64), consider using online platforms like DeepNote.com.

Structure

Comment Analysis Functions

train(train_csv, test_size=0.2, vector_size=100, window=5)

Parameter	Data Type	Default Value	Description	Optional/Required
train_csv	str	-	Path to CSV file containing training data with body and recommendation_status columns	Required
test_size	float	0.2	Proportion of test data (between 0.0 and 1.0)	Optional
vector_size	int	100	Output vector dimension for Word2Vec model (embedding size)	Optional
window	int	5	Context window size for Word2Vec model	Optional

recommendation_status must be one of:

• no_idea

• recommended

• not_recommended

Null/NaN values are converted to no_idea, affecting model accuracy.

• Returns test accuracy score.

---

analyzeText(text)

Parameter	Data Type	Description	Optional/Required
text	str	The Persian text to be analyzed	Required

The core function that analyzes a text and returns one of: "not_recommended", "recommended", or "no_idea".

---

saveModel()

loadModel()

Model persistence functions. Models are saved in the model directory.

---

analyzeCSV(input_csv, output_path, summary_path=None, text_column=0)

Parameter	Data Type	Default Value	Description	Optional/Required
input_csv	str	-	Path to input CSV file containing comments to analyze	Required
output_path	str	-	Path where analyzed results CSV will be saved	Required
summary_path	str or None	None	Optional path to save summary statistics CSV	Optional
text_column	int or str	0	Column index (int) or name (str) containing the text to analyze	Optional

Batch processes comments from a CSV file. For single-column files, text_column isn't needed. Otherwise specify column name/index (0-based, negative indices supported). Output contains:

1- Original text

2- Recommendation status Optional summary_path generates statistics:

• Total count

• Recommended count

• Not recommended count

• No idea count

• Model accuracy (not implemented in current version)

Returns a DataFrame and saves results.

---

Emotion Analysis Functions

loadLex(csv_file, word_col=0, emotion_col=1, weight_col=2)

Parameter	Data Type	Default Value	Description	Optional/Required
csv_file	str	-	Path to CSV lexicon file	Required
word_col	int or str	0	Column index (int) or name (str) containing words	Optional
emotion_col	int or str	1	Column index (int) or name (str) containing emotion labels	Optional
weight_col	int or str	2	Column index (int) or name (str) containing weight values	Optional

Loads a CSV with three columns:

1- Keywords

2- Emotion (happiness, sadness, anger, fear, disgust, calmness)

3- Emotion weight (defaults to 1 if unspecified, affecting accuracy)

Column indices are optional.

---

train(train_csv, text_col='text', emotion_col='sentiment', weight_col='weight')

Parameter	Data Type	Default Value	Description	Optional/Required
train_csv	str	-	Path to training CSV file	Required
text_col	str or int	'text'	Column name/index containing text data	Optional
emotion_col	str or int	'emotion'	Column name/index containing emotion labels	Optional
weight_col	str or int	'weight'	Column name/index containing weight values	Optional

Trains the emotion model using a CSV with specified column names (optional).

---

saveModel(model_name='weighted_sentiment_model')

Parameter	Type	Default Value	Description	Optional/Required
model_name	str	'weighted_sentiment_model'	Base filename for saving model (without extension)	Optional

loadModel(model_name='weighted_sentiment_model')

Parameter	Type	Default Value	Description	Optional/Required
model_name	str	'weighted_sentiment_model'	Base filename of model to load (without extension)	Optional

Model persistence functions (saved in model directory).

---

analyzeText(text)

Parameter	Type	Description	Optional/Required
text	str	Persian text to analyze	Required

Analyzes a single text, returning percentage scores for each emotion.

---

analyzeCSV(input_csv, output_csv, text_col='text', output_col='sentiment_analysis')

Parameter	Type	Default Value	Description	Optional/Required
input_csv	str	-	Path to input CSV file containing text to analyze	Required
output_csv	str	-	Path to save analyzed results	Required
text_col	str/int	'text'	Column name/index containing text to analyze	Optional
output_col	str	'sentiment_analysis'	Column name for output results	Optional

Batch processes texts from CSV. Returns True on success. Requires:

• input_csv path

• output_csv path Optional column names.

---

Installation

Install via pip:

pip install PerSent

For specific versions:

pip install PerSent==<VERSION_NUMBER>

Usage

• Comment Analysis

Basic single-text analysis:

from PerSent import CommentAnalyzer

analyzer = CommentAnalyzer()

'''
Training (if you have data):
Requires CSV with comments and recommendation status columns
Status must be: recommended/not_recommended/no_idea
'''
analyzer.train("train.csv")

# Load pre-trained model
analyzer.loadModel()

# Predict
text = "کیفیت عالی داشت" # "Excellent quality"
result = analyzer.analyzeText(text)
print(f"Sentiment: {result}")  # Output: Sentiment: recommended

The included pre-trained model has ~70% accuracy. For better results, you can train with larger datasets. I've prepared a split dataset (due to size):

Download Here

---

Batch CSV processing:

from PerSent import CommentAnalyzer
analyzer = CommentAnalyzer()
analyzer.loadModel()

# Basic usage (single-column CSV)
analyzer.analyzeCSV(
    input_csv="comments.csv",
    output_path="results.csv"
)

# Alternative usage patterns:
# 1. Using column index (0-based)
analyzer.analyzeCSV("comments.csv", "results.csv", None, 0)

# 2. Negative indices (count from end)
analyzer.analyzeCSV("comments.csv", "results.csv", None, -1)

# 3. Column name
analyzer.analyzeCSV("comments.csv", "results.csv", None, "نظرات") # "Comments" column

# 4. With summary (single-column)
analyzer.analyzeCSV("comments.csv", "results.csv", "summary.csv")

# 5. With summary and column specification
analyzer.analyzeCSV("comments.csv", "results.csv", "summary.csv", 2)

• Emotion Analysis

Single text analysis with pre-trained model:

from PerSent import SentimentAnalyzer

analyzer = SentimentAnalyzer()
analyzer.loadModel()

sample_text = "امتحانم رو خراب کردم. احساس می‌کنم یک شکست خورده‌ی تمام عیارم."
# "I failed my exam. I feel like a complete failure."

result = analyzer.analyzeText(sample_text)
for emotion, score in sorted(result.items(), key=lambda x: x[1], reverse=True):
    print(f"{emotion}: {score:.2f}%")

output :

غم: 36.00%                     #Sadness
عصبانیت: 36.00%                 #anger
ترس: 28.00%                    #fear
شادی: 0.00%                     #happiness
تنفر: 0.00%                      #disgust
شگفتی: 0.00%                    #surprise
آرامش: 0.00%                    #calmness

To train your own model:

analyzer.train('emotion_dataset.csv')

Required CSV columns:

1- Keywords

2- Emotion (happiness, sadness, anger, disgust, fear, calmness)

3- Emotion weight

Model persistence:

analyzer.saveModel("custom_model_name")
analyzer.loadModel("custom_model_name")

Batch CSV processing:

analyzer.analyzeCSV("input.csv", "output.csv")

Contribution

As mentioned, this library needs community collaboration. Please share suggestions, bugs, or feedback via:

• Fork Repository & Pull Request

• Create Issue

• Email: [e-mail]

• Telegram: @RezaGooner