نور سعدي الجبوري

This study investigated the effectiveness of conventional and deep learning models in recognizing the Iraqi Arabic dialect from text data. Accurately identifying regional dialects has become more important due to the spread of digital platforms. Its relevance can be seen in multiple domains, such as sentiment analysis, social media monitoring, and cultural preservation. In this work, a dataset was built where samples were collected from social media platforms, and are prepared and preprocessed to facilitate model training to classify three Iraqi dialects. The work implemented different approaches to classify the Iraqi dialects using conventional, machine learning, deep learning, Transformers, and hybrid models which combined different approaches. The models’ hyperparameters were tuned to achieve the best accuracy and performance in the eva‎luation of unseen datasets. The study investigated the performance of TF-IDF SVM, dictionary-based model, LSTM-based model, Hybrid LSTM model, and BERT-based transformers pre-trained on the general Arabic language. The hybrid model and the transformers model results exceeded the performance of the pure machine learning, deep learning, and dictionary-based models. The hybrid model combined the strength of both to achieve higher accuracy and f1-score. The transformers model utilized their strength in dealing with sequential input through attention and masking mechanisms where they achieved high accuracy similar to the hybrid model. The pure deep learning model did not achieve the highest accuracy due to the large similarities between the dialects and the existence of outliers in the dataset. The dictionary-based achieved low accuracy because it could not capture the sequential context of words used in each dialect.

يادگيري ماشيني، يادگيري عميق، LSTM، ترانسفورماتورها، گويش، زبان، طبقه بندي، LLM.

Machine Learning, Deep Learning, LSTM, Transformers, Dialect, Language, Classification, LLM.

noora aljubouri

Dr. Naderi hassan