Yassine-J/tunisian-sentiment-analysis-demo

A demo for a Tunisian sentiment analysis with a pre-trained model # Tunisian Sentiment Analysis with Arabic-BERT 🇹🇳 This project uses ARABERT, a BERT-based model pre-trained on Arabic dialects including Tunisian, to perform sentiment analysis on Tunisian Arabic text. ## Project Goals - Demonstrate Tunisian dialect sentiment analysis using transformer models - Provide an end-to-end pipeline from raw data to predictions - Showcase transfer learning with Arabic-BERT for low-resource dialects ## What is Arabic-BERT? Arabic-BERT is a family of BERT models specifically trained on Arabic text, including dialectal content. The base model has: - 12 layers - 768 hidden dimension - 110M parameters - Trained on 8.5B tokens of Arabic text Available sizes: `mini (41M)` | `medium (110M)` | `base (110M)` | `large (335M)` **Resources**: - Model Card - GitHub ## Dataset The dataset `sample.csv` contains Tunisian Arabic text with binary sentiment labels: | Column | Description | |--------|----------------------| | text | Tunisian Arabic text | | label | 0=Negative, 1=Positive | **Note**: if the csv file doesn't show the arabic words, do Data -> From Text/CSV -> UTF-8 -> Load **Example Data**: ```csv text,label "ما عجبنيش الخدمة",0 "خدمتهم زينة والله",1 ``` ## Quick Start ```python pip install transformers torch pandas ``` ```python from transformers import AutoTokenizer, AutoModelForSequenceClassification # Load model (replace with your fine-tuned path) model_name = "asafaya/bert-base-arabic" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=2) # Predict example inputs = tokenizer("ما نصحكش بهذا المطعم", return_tensors="pt") outputs = model(**inputs) ```