💬 Customer Support Tone Checker (Persian)

A small machine learning project for detecting the tone of Persian customer support messages.

The model classifies each message into one of three tone categories:

• polite
• semi_polite
• impolite

This project was built as a university project and as practice for working with Persian text preprocessing, TF-IDF features, machine learning classifiers, and a simple Gradio interface.

📌 Project Overview

Customer support messages can have different tones. Some messages are polite, some are neutral or semi-polite, and some may sound impolite.

The goal of this project is to build a simple text classification pipeline that can analyze Persian customer support messages and predict their tone.

This project includes:

• Persian text preprocessing
• TF-IDF feature extraction
• Training and comparing multiple machine learning models
• Model evaluation
• A simple Gradio interface for testing messages

---

📁 Project Structure

CustomerSupportToneChecker/

• dataset/

  • tone_dataset.csv

• models/

  • tfidf_vectorizer.pkl
  • svm_linearsvc.pkl
  • logistic_regression.pkl
  • random_forest.pkl
  • decision_tree.pkl

• notebooks/

  • 01_build_dataset.ipynb
  • 02_ToneDetection_api_ui.ipynb

• README.md

---

🗂️ Dataset

The dataset is located in:

dataset/tone_dataset.csv

It contains 300 Persian customer support messages.

The messages are divided into three labels:

• polite: polite customer support message
• semi_polite: partly polite or neutral message
• impolite: impolite customer support message

Dataset columns:

• id: unique message ID
• text: Persian customer support message
• label: tone category

---

🧹 Text Preprocessing

The preprocessing step includes:

• Normalizing Persian text
• Removing non-Persian characters
• Removing digits
• Tokenizing text
• Removing stopwords
• Removing very short tokens

I did not use stemming or lemmatization in this project because they can sometimes change the structure or meaning of Persian words.

🔢 Feature Extraction

The project uses TF-IDF to convert Persian text into numerical features.

Main TF-IDF settings:

• ngram_range=(1, 2)
• min_df=2
• max_features=5000
• sublinear_tf=True

The saved TF-IDF vectorizer is stored in:

models/tfidf_vectorizer.pkl

🤖 Machine Learning Models

The following machine learning models were trained and compared:

• SVM / LinearSVC
• Logistic Regression
• Random Forest
• Decision Tree

Performance summary:

• SVM / LinearSVC: Accuracy around 0.93, Macro F1 around 0.93
• Logistic Regression: Accuracy around 0.93, Macro F1 around 0.93
• Random Forest: Accuracy around 0.88
• Decision Tree: Accuracy around 0.80

SVM and Logistic Regression performed best in this project.

🖥️ Gradio Interface

This project includes a simple Gradio interface for testing Persian customer support messages.

The interface supports:

• Single message prediction
• Model selection
• Comparing model outputs
• Sample Persian messages
• Right-to-left Persian text display
• Optional Hugging Face API integration

The Gradio interface is available inside:

notebooks/02_ToneDetection_api_ui.ipynb

---

🚀 How to Run

1. Clone the repository

git clone https://github.com/fatsed/CustomerSupportToneChecker.git

2. Go to the project folder

cd CustomerSupportToneChecker

3. Install the required libraries

pip install pandas scikit-learn hazm gradio joblib

4. Open the notebook

Open this notebook:

notebooks/02_ToneDetection_api_ui.ipynb

Then run the cells to load the models and test the Gradio interface.

---

✨ Example Use

You can enter a Persian customer support message, and the model will predict whether the tone is:

• polite
• semi-polite
• impolite

This can be useful as a simple practice project for learning text classification and basic NLP concepts.

⚠️ Notes

This is a small educational project and not a production-level system.

Some limitations:

• The dataset is small.
• The dataset was created for practice.
• The semi_polite class can sometimes overlap with both polite and impolite.
• The model may not generalize well to real-world customer support data without a larger and more diverse dataset.

---

📚 What I Learned

Through this project, I practiced:

• Working with Persian text data
• Building a simple NLP pipeline
• Using TF-IDF for text classification
• Training different machine learning models
• Comparing model performance
• Saving and loading trained models
• Creating a simple Gradio demo

---

🛠️ Tech Stack

• Python
• Pandas
• Scikit-learn
• Hazm
• Gradio
• Joblib
• Jupyter Notebook