ArBanking77: Intent Detection Neural Model and a New Dataset in Modern and Dialectical Arabic

ArBanking77 is an MSA and Dialectal Arabic Corpus for Arabic Intent Detection in the Banking Domain. It consists of 31,404 samples (MSA, Palestinian, Saudi, Moroccan, and Tunisian dialects). This repo contains the source-code and dataset to train and evaluate Arabic Intent Detection model.

ArBanking77 Corpus

ArBanking77 consists of 31,404 (MSA, Palestinian, Saudi, Moroccan, and Tunisian dialects) that are manually Arabized and localized from the original English Banking77 dataset; which consists of 13,083 queries. Each query is classified into one of the 77 classes ( intents) including card arrival, card linking, exchange rate, and automatic top-up. You can find the list of these 77 intents in the ./data/Banking77_intents.csv file. A neural model based on AraBERT was fine-tuned on the ArBanking77 dataset (F1-score 92% for MSA, 90% for PAL)

Full Corpus Download

Data is available in the data directory for academic and commercial use. However, we cannot provide the augmented data.

Model Download

SinaLab HuggingFace

Online Demo

You can try our model using this demo link.

Requirements

At this point, the code is compatible with Python 3.11

Clone this repo

git clone https://github.com/SinaLab/ArBanking77.git

This package has dependencies on multiple Python packages. It is recommended that Conda be used to create a new environment that mimics the same environment the model was trained in. Provided in this repo requirements.txt from which you can create a new conda environment using the command below.

conda create -n env_name python=3.11

Install requirements using pip command:

pip install -r requirements.txt

Project Structure

.
├── data                            <- data dir
│   ├── Banking77_Arabized_MSA_PAL_train.csv
│   ├── Banking77_Arabized_MSA_PAL_val.csv
│   ├── Banking77_Arabized_MSA_test.csv
│   ├── Banking77_Arabized_PAL_test.csv
│   ├── Banking77_Arabized_Moroccan_test.csv
│   ├── Banking77_Arabized_Saudi_test.csv
│   ├── Banking77_Arabized_Tunisian_test.csv
│   ├── Banking77_intents.csv
├── outputs
│   ├── models                      <- trained models
│   ├── results                     <- evaluation results and reports
├── src                             <- training and evaluation scripts
│   ├── run_glue_no_trainer.py
│   ├── run_glue_no_trainer_eval.py
│   └── utils.py
├── .gitignore
├── LICENSE
├── README.md
└── requirements.txt

Model Training

You can start model training by running the following command. It's recommended to pass the arguments demonstrated below to get results similar to the ones reported in the paper.

python ./src/run_glue_no_trainer.py
    --model_name_or_path aubmindlab/bert-base-arabertv02 
    --train_file ./data/Banking77_Arabized_MSA_PAL_train.csv
    --validation_file ./data/Banking77_Arabized_MSA_PAL_val.csv 
    --seed 42 
    --max_length 128 
    --learning_rate 4e-5 
    --num_train_epochs 20 
    --per_device_train_batch_size 64 
    --output_dir ./outputs/models

Evaluation

Additionally, you can evaluate the trained model on Banking77_Arabized_MSA_test.csv, Banking77_Arabized_PAL_test.csv, Banking77_Arabized_Moroccan_test.csv, Banking77_Arabized_Saudi_test.csv, and Banking77_Arabized_Tunisian_test.csv test sets as follows:

python ./src/run_glue_no_trainer_eval.py 
    --model_name_or_path ./outputs/models 
    --validation_file ./data/Banking77_Arabized_MSA_test.csv 
    --seed 42 
    --per_device_eval_batch_size 64 
    --results_dir ./outputs/results 
    --log_path ./outputs/logs/log.txt

Credits

The first phase of this research was partially funded by the Palestinian Higher Council for Innovation and Excellence and the Scientific and TÜBİTAK under project No. 120N761 - CONVERSER: Conversational AI System for Arabic.

Citation

Mustafa Jarrar, Ahmet Birim, Mohammed Khalilia, Mustafa Erden, and Sana Ghanem: ArBanking77: Intent Detection Neural Model and a New Dataset in Modern and Dialectical Arabic. In Proceedings of the 1st Arabic Natural Language Processing Conference (ArabicNLP), Part of the EMNLP 2023. ACL.

Sanad Malaysha, Mo El-Haj, Saad Ezzini, Mohammed Khalilia, Mustafa Jarrar, Sultan Nasser, Ismail Berrada, Houda Bouamor: AraFinNLP 2024: The First Arabic Financial NLP Shared Task. In Proceedings of the Second Arabic Natural Language Processing Conference (ArabicNLP 2024), Bangkok, Thailand. Association for Computational Linguistics.