| | --- |
| | library_name: sklearn |
| | tags: |
| | - text-classification |
| | - persian |
| | - offensive-language-detection |
| | - onnx |
| | language: fa |
| | metrics: |
| | - accuracy |
| | - f1 |
| | - precision |
| | - recall |
| | license: mit |
| | --- |
| | |
| | # Offensive Text Classifier for Persian Language |
| |
|
| | This repository contains an Offensive Text Classifier model trained to detect offensive language in Persian text. The model is built using `scikit-learn` and exported to ONNX format for efficient inference. |
| |
|
| | ## Model Description |
| |
|
| | This model classifies Persian text into two categories: `Offensive` (label 1) and `Neutral` (label 0). It uses a TfidfVectorizer for text feature extraction combined with a Support Vector Machine (SVM) classifier. |
| |
|
| | ## Dataset |
| |
|
| | The model was trained on the [ParsOffensive dataset](https://github.com/golnaz76gh/pars-offensive-dataset). This dataset consists of Persian comments labeled as either 'Offensive' or 'Neutral'. |
| |
|
| | ## Preprocessing |
| |
|
| | The text data underwent the following preprocessing steps: |
| | - **Normalization**: Using `hazm.Normalizer`. |
| | - **Lemmatization**: Using `hazm.Lemmatizer`. |
| | - **Stop-word Removal**: Common Persian stop words were removed. |
| | - **Label Encoding**: 'Neutral' and 'Offensive' labels were converted to numerical `0` and `1` respectively. |
| | - **Imbalance Handling**: The ADASYN technique was applied to address class imbalance during training. |
| |
|
| | ## Model Architecture |
| |
|
| | The final pipeline consists of: |
| | 1. **`TfidfVectorizer`**: Converts raw text into a matrix of TF-IDF features. |
| | 2. **`SVC` (Support Vector Classifier)**: A Support Vector Machine classifier with a radial basis function (RBF) kernel. |
| |
|
| | ## Performance |
| |
|
| | Below are the performance metrics on the test set: |
| |
|
| | ``` |
| | precision recall f1-score support |
| | |
| | 0 0.80 0.96 0.88 1043 |
| | 1 0.91 0.62 0.74 644 |
| | |
| | accuracy 0.83 1687 |
| | macro avg 0.86 0.79 0.81 1687 |
| | weighted avg 0.84 0.83 0.82 1687 |
| | ``` |
| |
|
| | Detailed Metrics: |
| | - **Accuracy**: 0.830 |
| | - **Precision (Offensive)**: 0.909 |
| | - **Recall (Offensive)**: 0.618 |
| | - **F1-score (Offensive)**: 0.736 |
| |
|
| | ## How to Use |
| |
|
| | ### Load the model |
| |
|
| | You can load the ONNX model and use it for inference. You will need to apply the same preprocessing steps as during training. |
| |
|
| | ```python |
| | from hazm import Lemmatizer, Normalizer, stopwords_list |
| | |
| | # recreate the preprocessing components (or load them if saved) |
| | stopwords = stopwords_list() |
| | lemmatizer = Lemmatizer() |
| | normalizer = Normalizer() |
| | |
| | def clean_sentences(sentence: str) -> str: |
| | return " ".join(lemmatizer.lemmatize(word) for word in normalizer.normalize(sentence).split(" ") if word not in stopwords) |
| | ``` |
| |
|
| | ## Dependencies |
| |
|
| | - `pandas` |
| | - `numpy` |
| | - `hazm` |
| | - `scikit-learn` |
| | - `imblearn` |
| | - `skl2onnx` |
| | - `onnxruntime` |
| | - `huggingface_hub` |
| |
|
| | ## Demo |
| | You can try it [here](https://huggingface.co/spaces/sirunchained/bad-comment-gradio). |
