This is a simple text classification library focused on dialect identification, based on keras. Some Arabic text normalization utilities are included.

1- Word Level CNN based on [Convultion Neural Network for Text Classificartion].

2- Word Level C-LSTM based on [A C-LSTM Neural Network for Text Classification].

3- Recurrent Network and its variants (BiLSTM, LSTM, GRU, BiGRU, Attention-BiLSTM)

4- Models implemented but currently not supported in options (Attention-LSTM,Attention-BiGRU).

5- Not yet tested (char level CNN).

- keras (2.0 or above)
- gensim
- numpy
- pandas

- * Tested with python 3.4 *
- python test_baselines.py --train training_file --Ar='True' --dev Dev_File --test test_file --model_type=model_selection --static=Trainable_embeddings --rand=Random_Embeddings --embedding=External_Embedding_model --model_file=Output_model_file_inJson
- put your training labels in [[link](https://github.com/UBC-NLP/aoc_id/edit/master/conf/label_list)].

• train: training file assuming in csv format, text, label
• Ar: if True then Arabic normalization is applied (should be true in case of external embeddings)
• dev: Development file in csv format
• test: test file in csv format
• model_type: currently support those type of models: (cnn: word level cnn, clstm: word level clstm, lstm: vanilla lstm architecture, blstm: Vanilla bidirectional LSTM, bigru: Vanilla BiGated Recurrent unit, attbilstm: BiLSTM with self attention mechanism)
• static: used in case of external embedding, if True: External Embeddings are not fine tuned during training, if False: External EMbeddings are fine tuned during training).
• rand: if True, No external embedding is applied, randomly initialized embedding
• embedding: External embedding model in gensim format
• model_file: Output model file in Json. -EMB_type: Choose whether fastText or CBOW or skipgram

Note: final model score is dumped into a file with name_of_model_score with both dev and test scores

• This project utilize 6 deep learning models applied on Arabic Online Commentary Dataset. [Paper]; [Dataset].

• Make sure to cite AOC oringial paper if you are going to use it in your work.

• This work currently accepted to VarDial Worshop 2018 co-located with COLING 2018 under the name (paper soon) "Deep Models for Arabic Dialect Identification on Benchmarked Data"

• Training data: [link]

• Dev data: [link]

• Test data: [link]

• An example on how to use it is in: [link]

• If you are going to follow up on this project please cite this work using the following bibtext:

@inproceedings{Elaraby2018,
  title={Deep Models for Arabic Dialect Identification on Benchmarked Data},
  author={Elaraby, Mohamed and Abdul-Mageed, Muhammad},
  booktitle={Proceedings of the Fifth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial5)},
  year={2018}
}

• For Arabic Dialects we release 2 embedding models
• AOC embedding: [Download URL]
• Twitter Embedding Model: [Download URL]
• cite the following paper if you are planning to use city level dialect embedding model:

@inproceedings{mageedYouTweet2018,
  title={You Tweet What You Speak: A City-Level Dataset of Arabic Dialects},
  author={Abdul-Mageed, Muhammad and Alhuzali, Hassan and Elaraby, Mohamed},
  booktitle={LREC},
  pages={3653--3659},
  year={2018}
}